TELEOSEMANTICS

This page intentionally left blank

Teleosemantics
New Philosophical Essays

Edited by
GRAHAM MACDONALD
AND DAVID PAPINEAU

CLARENDON PRESS · OXFORD

 1
Great Clarendon Street, Oxford  
Oxford University Press is a department of the University of Oxford.
It furthers the University’s objective of excellence in research, scholarship,
and education by publishing worldwide in
Oxford New York
Auckland Cape Town Dar es Salaam Hong Kong Karachi
Kuala Lumpur Madrid Melbourne Mexico City Nairobi
New Delhi Shanghai Taipei Toronto
With offices in
Argentina Austria Brazil Chile Czech Republic France Greece
Guatemala Hungary Italy Japan Poland Portugal Singapore
South Korea Switzerland Thailand Turkey Ukraine Vietnam
Oxford is a registered trademark of Oxford University Press
in the UK and in certain other countries
Published in the United States
by Oxford University Press Inc., New York
 the several contributors 2006
The moral rights of the authors have been asserted
Database right Oxford University Press (maker)
First published 2006
All rights reserved. No part of this publication may be reproduced,
stored in a retrieval system, or transmitted, in any form or by any means,
without the prior permission in writing of Oxford University Press,
or as expressly permitted by law, or under terms agreed with the appropriate
reprographics rights organization. Enquiries concerning reproduction
outside the scope of the above should be sent to the Rights Department,
Oxford University Press, at the address above
You must not circulate this book in any other binding or cover
and you must impose the same condition on any acquirer
British Library Cataloguing in Publication Data
Data available
Library of Congress Cataloging in Publication Data
Data available
Typeset by Laserwords Private Limited, Chennai, India
Printed in Great Britain
on acid-free paper by
Biddles Ltd., King’s Lynn, Norfolk

ISBN 0–19–927026–0 978–0–19–927026–2

ISBN 0–19–927027–9 (Pbk.) 978–0–19–927027–9 (Pbk.)

1 3 5 7 9 10 8 6 4 2
 Contents

List of Contributors vii

Introduction: Prospects and Problems for Teleosemantics 1

Graham Macdonald and David Papineau
1. Language, Modularity, and Evolution 23
Kim Sterelny
2. Mental Representation, Naturalism, and Teleosemantics 42
Peter Godfrey-Smith
3. Representation, Teleosemantics, and the Problem of
Self-Knowledge 69
Fred Dretske
4. The Epistemological Objection to Opaque Teleological
Theories of Content 85
Frank Jackson
5. Useless Content 100
Ruth Millikan
6. On Thinking of Kinds: A Neuroscientific Perspective 115
Dan Ryder
7. Teleosemantics and the Consumer 146
Mohan Matthen
8. Content for Cognitive Science 167
Karen Neander
9. Representation and Unexploited Content 195
Robert Cummins, Jim Blackmon, David Byrd,
Alexa Lee, and Martin Roth
10. Fearing Fluffy: The Content of an Emotional Appraisal 208
Carolyn Price

Index 229
This page intentionally left blank
 List of Contributors

Jim Blackmon is a doctoral student in the Philosophy Department at the Univer-

sity of California at Davis.
David Byrd is a Lecturer in the Philosophy Department at Santa Clara University.
Robert Cummins is Professor in the Philosophy Department at the University of
Illinois-Urbana/Champaign.
Fred Dretske is Emeritus Professor of Philosophy at Stanford University.
Peter Godfrey-Smith is Professor in the Philosophy Department at Harvard Uni-
versity.
Frank Jackson is Distinguished Professor of Philosophy and Director of the Re-
search School of Social Sciences, the Australian National University.
Alexa Lee is a doctoral student in the Philosophy Department at the University of
California at Davis.
Graham Macdonald is Professor of Philosophy in the School of Philosophy
and Religious Studies, University of Canterbury, New Zealand, and Distin-
guished International Fellow, Institute of Cognition and Culture, Queen’s
University, Belfast.
Mohan Matthen is Canada Research Chair in Philosophy, Perception, and Com-
munication at the University of Toronto.
Ruth Millikan is Board of Trustees Distinguished Professor Emerita in the Philo-
sophy Department at the University of Connecticut.
Karen Neander is Professor of Philosophy at the University of California at Davis.
David Papineau is Professor of Philosophy of Science at King’s College London.
Carolyn Price is Lecturer in Philosophy at the Open University.
Martin Roth is an Assistant Professor of Philosophy at Knox College, Illinois.
Dan Ryder is an Assistant Professor in the Department of Philosophy at the Uni-
versity of Connecticut.
Kim Sterelny is a Professor in the Philosophy Department at the Australian
National University and at the Victoria University of Wellington, New Zealand.
This page intentionally left blank
 Introduction: Prospects and Problems
for Teleosemantics
Graham Macdonald and David Papineau

1. NATURALISM

The programme that has come to be known as ‘teleosemantics’ aims to offer a

naturalistic account of mental representation. That is, it aims to show how
the representational powers of mental states fit into the world revealed by
the natural sciences. It is distinguished from other naturalistic approaches
to mental representation by its reliance on a notion of function that plays
a prominent role in biology: thus it explains the truth conditions of belief-
like states, say, or the satisfaction conditions of desire-like states, in terms
of the biological functions of these states. In the next section we shall give
fuller details of the ways in which teleosemanticists have used the notion of
function to explain representation. But first it is worth considering exactly
how teleosemantics might contribute to the general project of naturalization.
The notion of ‘naturalism’ is much contested. The requirement that we
should only credit those facts that are recognized by the ‘natural sciences’
has little bite in the absence of some substantial specification of what quali-
fies a domain of investigation as a ‘natural science’. For many contemporary
philosophers, the real significance of naturalism lies in the need to explain
what role, if any, mental and other prima facie non-physical phenomena
play in the physical world. The rationale for this emphasis is that physics,
unlike other natural sciences, seems to be causally complete: physical effects
always have physical causes, if they have causes at all. This means that any
facts that are capable of producing physical effects cannot be ontologically
supplementary to the physical realm; otherwise there would be an absurd
proliferation of causal overdetermination. For example, since mental facts
often have physical effects in the form of bodily movements and consequent
2 Graham Macdonald and David Papineau

impacts on the wider world, this argues that they cannot themselves lie out-
side the physical realm, for if they did then the bodily movements and so on
would have non-physical causes as well as physical ones.
Still, the doctrine that the mental realm cannot be outside the physic-
al realm is less straightforward than it seems. For one thing, it is unclear
how ‘physical’ is to be understood in this context. As Carl Hempel poin-
ted out some decades ago (Hempel 1969), our present best physical theory
is likely to be overtaken in the future by a better theory, which argues
that we would be wrong to constrain our naturalizing ambitions by our
present understanding of the physical. On the other hand, the constraint
that our ontology must respect some future physical theory seems no con-
straint at all, given our ignorance of that future theory. So any attempt
properly to articulate the idea that the mental cannot be separate from
the ‘physical’ threatens to make this doctrine either overly restrictive or
quite empty.
It seems to us that this challenge can be satisfactorily answered. The
recent literature contains a number of alternative suggestions for defining
‘physical’ in a way which leaves physicalism both plausible and contentful
(for a brief survey of the options, see Jackson 1998: 6–9).
However, even given such a definition of ‘physical’, there is another
familiar respect in which physicalism about the mental needs further elu-
cidation: how strictly should we read the requirement that the mental not
be ‘ontologically supplementary’ to the physical realm, as we put it above?
Few naturalists would want to subscribe to ‘type physicalism’, in the sense
of requiring each respectable mental property to be strictly identical with
some property describable in the language of physics. Rather, ontological
indistinctness from the physical is widely agreed by physicalists to require
only that mental properties should supervene on physical facts, in the sense
that they are metaphysically determined by the physical facts. But now this
threatens to remove the teeth from naturalism once more. Supervenience
on the physical is not a strong requirement: for example, moral properties
are generally supposed to supervene on physical properties, even by philo-
sophers who would strongly resist the idea that moral facts can somehow be
investigated by the methods of the natural sciences.
Still, the demands of supervenience are not empty. If you hold that
certain properties, while not type-identical to physical properties, are nev-
ertheless metaphysically supervenient on them, then surely you owe some
explanation of why this should be so. What is it about mental properties,
say—or indeed moral properties—that makes it the case that a mental or
moral difference must be due to a physical difference? A satisfactory answer
need not type identify mental or moral properties with physical properties,
but it will need to give some account of the nature of these properties that
 Prospects and Problems for Teleosemantics 3

will explain why their instances should be metaphysically fixed by the phys-
ical facts.
The view that representational facts are functional facts can be seen as an
answer to this challenge. The concept of function that is used in biology is
itself a contested notion. In fact, it is likely that there is no ‘one’ notion of
function employed in biology. We shall consider some alternative analyses
of ‘function’ below. But, on any account, two things are clear. First, func-
tional properties are a paradigm of properties that are not type-reducible to
physical properties. There is no strictly physical property that is necessary
and sufficient for being a wing, say. All it takes for something to be a wing is
that it have the function of enabling flight: beyond this there is no limit to
the physical variety of different kinds of wing. And the same goes for other
familiar functional categories in biology, like being a stomach, or a heart, or
an eye. Second, despite this lack of type reducibility, it is clear that the func-
tional facts are metaphysically determined by the physical facts. Two items
could not possibly have all the same physical features (including their phys-
ical histories and environments) yet not have the same functional features.
Once the (wide) physical properties of something are given, then its func-
tional nature is fixed. (Different analyses of biological function will explain
this supervenience on the physical in different ways, but all will agree that
the functional facts do supervene on the physical facts.)
So an analysis of representational facts as functional facts will imply that
representational properties are not type-identical to physical properties, yet
at the same time will explain why representational facts must supervene on
the physical facts and thus be naturalistically acceptable.
Even if teleosemantics does not type reduce representational properties
to physical properties, for the reasons just explained, it may reduce them to
biological properties. (Thus Ruth Millikan, the most prominent of teleose-
manticists, entitled her first book Language, Thought, and Other Biological
Categories, 1984.) In defence of this biologically reductionist view, it can be
observed that teleosemantics aims to offer an explicit account of representa-
tional properties by appealing to a notion of function that is used in biolo-
gical theorizing. On the other hand, it is unclear whether the facts to which
teleosemanticists reduce representational properties should really qualify as
biological facts, given that they standardly involve cognitive mechanisms
that would normally be counted as in the realm of psychology rather than
biology. In the end, we do not think that much hangs on whether we think
of teleosemantics as a type reduction to biological properties. The more
important point is that teleosemantics offers a naturalistically acceptable
explanation of representation, whether or not we also count this as a bio-
logical reduction.
4 Graham Macdonald and David Papineau

2. THE TELEOSEMANTIC PROGRAMME

In this section we will explain the basic strategy of teleosemantics and

give some indication of the different ways this strategy has been developed
by different theorists. The simplest way to introduce the teleosemantic
programme is to contrast it with an alternative naturalistic approach to
mental representation, namely, causal or indicator semantics. On the latter
view, the content of a belief-like cognitive state is that condition that
typically causes the state, and which the state therefore indicates (Stampe
1977; Dretske 1981). The standard objection to this causal approach is that
it has trouble explaining misrepresentation. Misrepresentation by a belief-
like state occurs when the state is tokened, but its truth condition is not.
However, if the state’s truth condition is simply the range of circumstances
that cause the state to be tokened, then it is unclear how the state can be
tokened and yet its truth condition not obtain. Take a state that intuitively
represents the presence of a snake. Such a state will often be caused, not by
real snakes, but also by glimpses of slithery animals, toy snakes, and so on.
The problem for the causal theory is that it has no obvious way of excluding
these misleading extra causes from this state’s truth condition. So the causal
theory seems to end up implying, absurdly, that all tokenings of this belief-
like state are true.
The teleosemantic approach, by contrast, explains the content of a belief-
like state, not in terms of its typical causes, but in terms of how it is bio-
logically designed to function. The snake-registering state will be designed
to prompt behaviour that is advantageous specifically in the presence of real
snakes, and not in the presence of harmless slithery animals, or toy snakes.
So the truth condition of the state is specifically snake, and accordingly it
will misrepresent the environment when it is prompted by other causes.
Let us spell this out in a little more detail, using some notions introduced
by Ruth Millikan (1984, 1993). She distinguishes the mechanisms that pro-
duce mental representations from those that consume them. The producing
mechanisms will be the sensory and other cerebral mechanisms that give
rise to cognitive representations. The consumer mechanisms will be those
that use these representations to direct behaviour in pursuit of some bio-
logical end. Now, biological functions are in the first instance always a
matter of effects: a trait’s function is that effect it is supposed to produce.
So the function of a mental representation will lie in the way it contributes
to the biological end of the mechanism that consumes it. More specific-
ally, its function will be to enable the consumer mechanism to achieve its
end by gearing behaviour to circumstances. Given this, we can think of the
 Prospects and Problems for Teleosemantics 5

representation’s truth condition as the circumstance that enables it to fulfil

this function—that is, the circumstance in which the behaviour it prompts
is designed to produce the consumer mechanism’s end. For example, if we
think of the snake representation above as being consumed by a mechan-
ism whose function is to avoid snake bites, then the representation’s truth
condition will be snake, rather than harmless look-alikes, since the snake-
avoidance behaviour has been designed to have a positive result specifically
when a real snake is present.
Viewed in this way, teleosemantics has close affinities with the ‘success
semantics’ analysis of content within the context of everyday belief–desire
psychology (Whyte 1990, 1991). According to success semantics, the truth
condition of a given belief is that circumstance which will ensure the sat-
isfaction of whichever desire combines with the belief to prompt action.
(More intuitively, what shows that you believe p is that you choose beha-
viour that will satisfy your desires if p. What shows you believe a snake is
present is that you act in way that is sensible if a snake is present.)
To see success semantics as a special case of Millikan’s version of tele-
osemantics, we need only equate the consumer mechanism for a belief with
the decision-making processes that use that belief to select behaviour that
will satisfy your currently active desires. Given this, the association of a
success condition with a belief can be viewed as one example of the way
Millikan’s analysis fixes the content of any belief-like representation as that
circumstance under which a consumer mechanism guided by that belief
will achieve its end.
Not everybody views success semantics this way. Most philosophers who
have developed success semantics view it, not as an aspect of teleosemantics,
but as an alternative to it. And certainly there is no immediate appeal to
biological function in the idea of a circumstance that ensures the satisfac-
tion of the desire that combines with some belief to prompt action. On
the other hand, the success semantics programme is in one sense obviously
incomplete as an explanation of mental representation: it explains truth
conditions for beliefs in terms of satisfaction condition for desires. A full
account of mental representation needs to set the assumptions of success
semantics in a wider context, so as to yield an explanation of desire satis-
faction as well. And one obvious option here is to appeal to the notion of
biological function, and say that the satisfaction condition of a desire is that
result which the desire has the biological function of producing.
The resulting combination of success semantics and teleosemantics has
been developed in some detail by David Papineau (1984, 1993). This kind
of approach is ‘top-down’—it starts with the kind of complex cognitive
structure assumed by everyday belief–desire psychology—by contrast with
6 Graham Macdonald and David Papineau

Millikan’s ‘bottom-up’ approach—which begins with primitive biologic-

al representations of danger and food in simple non-human animals. In
favour of Millikan’s strategy, there is the obvious advantage of more general
applicability, and moreover her approach avoids the danger that everyday
belief-psychology may offer a misleading picture of actual human cognitive
structure. On the other hand, a full account of mental representation will
need to cover human cognition too, and Papineau’s approach offers one
possible account of this. In the end, perhaps the two approaches are best
thought of as complementary rather than competing.
Rather than trying to adjudicate between these alternatives, let us focus
instead on one feature they have in common, and which differentiates them
strikingly from causal or indicator semantics, and indeed from nearly all tra-
ditional philosophical analyses of content. Note that, on either account, the
processes giving rise to representations (the producing mechanisms) play no
particular role in the above analysis of content. On traditional approaches,
by contrast, the content of a representation is taken to be some kind of
function of the conditions that give rise to the representation, or correctly
give rise to it, or verify it, or some such. These approaches thus make it
relatively difficult, so to speak, for a representation to be false. (Of course,
they aim to avoid the charge, levelled at indicator semantics above, that they
make falsity impossible; but, even so, their general tenor is to make falsity
unusual.) By contrast, the teleosemantic approach as explained above disso-
ciates the determination of content from input conditions, and correspond-
ingly makes it very easy for representations to be false. On the teleosemantic
approach, content depends on how consumer mechanisms interpret rep-
resentations. It depends on the behavioural output, not the informational
input. The content is that condition under which the resulting behaviour
would be appropriate, whether or not the actual circumstances that caused
the representation are of that type.
Take the snake representation again. This is a snake representation
because it makes you behave in a way appropriate to snakes, given your
biological ends. And this will remain the case even if you are pretty bad
at recognizing snakes. The production mechanism for this representation
may be triggered by toy snakes, by other slithery animals, indeed by the
slightest hint of a slither, yet the representation will still stand for snake, if it
is specifically snake-appropriate behaviour that it prompts.
True, given that this representation has the content snake, in virtue of
this being the condition under which it has functional effects, its producing
mechanism will derivatively have the function of producing this representa-
tion only when it is true, that is, when snakes are present. Still, the fact that
this mechanism has this function by no means implies that it will achieve it
particularly often. To cite an oft-repeated example, sperm have the function
 Prospects and Problems for Teleosemantics 7

of fertilizing ova, but only one in a zillion actually does this. Provided the
pay-off from success sufficiently outweighs the costs of failure, biological
mechanisms can have functions that they fail to achieve far more often than
not. Indeed there are good reasons why we should expect a snake repres-
entation production mechanism to have just this structure: the pay-off from
success (avoiding a real danger of snake bite) so far outweighs the costs of
representing falsely (needless evasive action) that it makes biological sense
to err on the side of caution, and produce the representation in response to
even the most fallible signs of snakes.
So teleosemantics, as so far outlined, stands diametrically opposed to the
kind of input-based causal or verificationist theories that imply that false rep-
resentations are atypical. Given the frequency with which false representa-
tions are in fact found, this would seem to count in favour of the teleosemant-
ic approach. However, not all thinkers within the teleosemantic camp regard
its commitment to output-based content as an unalloyed advantage. Con-
sider the following well-known thought-experiment devised by Paul Piet-
roski (1992), and discussed by a number of contributors to this volume. The
kimu are simple creatures, with very limited sensory abilities, whose only
enemies are the snorf, who hunt them every day at dawn. A mutation endows
one of them with a disposition to sense and approach red things. This dispos-
ition is a biological advantage to its possessors, since it leads them to climb a
nearby hill every dawn, the better to observe the red sunrise, and means that
they thereby avoid the marauding snorf, who do not climb hills. As a result,
the disposition spreads through the kimu population.
Now, consider the state a kimu gets into when it is stimulated by some-
thing red. It seems natural to credit this state with the content red. But
an output-based teleosemantics argues differently. Nothing good happens
to the kimu just because they approach something red. Most of their red-
approaching behaviour is just a waste of time. It is only when this behaviour
takes them away from the dangerous snorf that it yields any biological
advantage. So an output-based teleosemantics will deem the state in ques-
tion to represent snorf-free, or predator-free, or some such.
This strikes many as strongly counter-intuitive, especially when it is fur-
ther specified that the kimu cannot tell a snorf from a sausage, and would
be perfectly happy to approach any snorf who happened to colour them-
selves red. Whatever the other virtues of teleosemantics, it seems wrong for
it to conclude that the kimu’s state signifies snorf-free, rather than simply
red. After all, by hypothesis the kimu’s senses are tracking the presence or
absence of redness, not the presence or absence of snorf.
There are alternative versions of teleosemantics that promise to analyse
cases like these differently. These alternatives place more emphasis on the
processes that produce representations than the purely output-based kind
8 Graham Macdonald and David Papineau

of teleosemantics described so far. For example, Fred Dretske builds his

version of teleosemantics on a prior notion of indication (1988, 1995).
Dretske first specifies that a type of state F indicates a type of state G just
in case Fs never occur in relevant environments in the absence of Gs. In
this sense, we can expect the products of sensory mechanisms to indic-
ate the stimuli that trigger those outputs, independently of any biological
advantages that may then ensue (and thus we can expect the kimu states
to indicate redness rather than snorf-freeness). Of course, indication in
itself does not amount to semantic representation, for the kind of reas-
ons given earlier: if the indicated condition is always present when the
indicator is, as the definition of indication requires, then the possibility of
misrepresentation has not yet been explained. So Dretske specifies that true
representation occurs only when F in addition has the function of indicat-
ing G—for example, the states of the visual system have the function of
indicating features of the nearby environment. Since items with biologic-
al functions can sometimes fail to perform these functions, so F can on
occasion fail to indicate G—this will happen when F is tokened in envir-
onments other than those where it has the function of being a sure-fire
indicator of G.
Dretske’s specific theory assumes a very strong notion of indication—Fs
never occur in relevant environments without Gs—and this generates par-
ticular problems for his approach (see McLaughlin 1991). Other philo-
sophers have focused on weaker notions of ‘indication’, requiring only that
Fs be correlated with Gs, not that they are sure-fire signs of Gs, while
still following Dretske’s suggestion that representation should be explained
as a matter of states having the function of indicating something, rather
the function of guiding behaviour towards some end. (Cf. Neander 1995,
Chapter 8 in this volume.)
However, whatever notions of indication they use, all such ‘input-based’
versions of teleosemantics face difficulties in explaining which correlations
between Fs and Gs count for representational purposes. There will be many
different kinds of Gs that any given type F is correlated with (even if we
require, with Dretske, that the correlation be perfect over certain environ-
ments): a given cortical sensory state, for instance, will co-vary with kinds of
surface stimulations of the sensory organs, and with intervening neural activ-
ations, as well as with a range of different distal external causes, such as slithery
appearances, actual snakes, forked tongues, and so on. True, all versions of
teleosemantics, including output-based ones, have some difficulty in explain-
ing how representations can be directed at some specific option from such a
range of alternatives (cf. Fodor 1990). However, output-based theories can
at least rule out candidate contents that do not ensure that advantageous con-
sequences will follow from resulting behaviour (such as slithery appearances,
 Prospects and Problems for Teleosemantics 9

as opposed to actual snakes). But input-based theories cannot appeal to this

resource, given that they aim to explain representational relations without
attending to behavioural consequences, and so face greater problems in deal-
ing with the threat of representational indeterminacy.
What about the counter-intuitive consequences that Pietroski’s
kimu thought-experiment seems to foist on output-based versions of
teleosemantics? Perhaps there is room for output-based teleosemantic
theories to argue that these intuitions depend on reading more into
Pietroski’s scenario than is justified by his description. Pietroski says that
the kimu evolve some state that is triggered by redness and which has the
advantage of keeping them away from the snorf. Given this specification,
it is hard to stop ourselves thinking of the kimu as having some general-
purpose visual system which gathers items of visual information which
might then be used to inform an open-ended range of behavioural projects
directed at different possible ends (such as avoiding blood, or finding
postboxes, or indeed wanting to see red things). However, this extra
structure in fact takes us significantly beyond what Pietroski’s description
actually requires, and it is open to output-based teleosemanticists to argue
that their theory is quite able to explain why an organism with all this
extra structure would be representing redness rather than snorf-freeness: if
the organism’s visual states inform a range of different behaviours directed
at different ends, then the content of any such state needs to be fixed as
some condition that assists in the achievement of all those ends, and this
may well come out as redness. On the other hand, if we do stick to a
minimal understanding of the snorf, as having only a special-purpose visual
sensitivity that brings no advantage except snorf-avoidance, then it’s not so
clear that there is anything wrong with the output-based reading of their
states as representing snorf-freeness: after all, if these states never do anything
except trigger simple avoidance behaviour, it seems natural enough to read
them as representing the danger they are designed to avoid.

3. FUNCTIONS

All teleosemanticists seek to analyse content in terms of biological function.

But as yet we have said nothing explicit about the notion of biological func-
tion itself. In fact, this notion is much disputed. The two main contenders
are the ‘historical–etiological’ view, and the ‘systems’ view that is analysed
and defended by Robert Cummins. As it happens, the former is favoured
by most teleosemanticists. But there is no necessary tie here, and in prin-
ciple one could combine a teleosemantic approach to representation with a
systems view of teleology.
10 Graham Macdonald and David Papineau

The systems account of functions focuses on complexity and aims to

understand the complex functioning of biological and other systems in
terms of the working of their parts (Cummins 1975, 2002; see also Mil-
likan 2002). Many complex systems can be thought of a goal-directed, and
in such cases the systems account of function offers one way of understand-
ing how biologists are able to explain traits by citing their effects: on the
systems account, such ‘functional explanations’ show how a part of a system
contributes to the system’s production of some goal.
Cummins distinguishes explanations of changes, which answer the ques-
tion ‘What caused system S to acquire property P?’, from explanations of
properties, which tell us ‘what it is for system S to instantiate property P’.
Property explanations can be given by constructing an analysis which details
the properties of S’s components and their organization. For example, the
kinetic theory of gases explains what it is for a gas to have temperature via
an account of the properties of the molecules contained within the gas.
Cummins calls the analysis of such a system ‘compositional analysis’. He
distinguishes ‘functional analysis’ as the analysis of dispositional properties,
like the ability to see, or digest, or locomote. So for Cummins the ascrip-
tion of a function to some part of a system specifies how the part has some
capacity that helps the system to achieve some result. In biology, organisms
are analysed into a number of systems (such as the digestive system, the
circulatory system, and so on) with each system having a particular task to
perform. How these tasks get performed is in turn explained by functional
analysis, such an analysis citing the capacities of parts of these systems, these
capacities being the functions of the parts. It will be a constraint on such
analyses that the properties of the parts of the system be less sophisticated,
less complex, than the property of the system being analysed.
By its nature, functional analysis is most fruitfully applied to complex
systems whose behaviour is produced by the coordinated action of simpler
but well-organized parts. It is unsurprising, then, that the paradigms of sys-
tems suitable for functional analysis are those that have been designed to
operate via the relevant interdependence of the parts. There is nothing in
the systems account of functions, however, that restricts it only to designed
systems: in principle any complex system can be subject to a systems-style
functional analysis, including such non-designed systems as the terrestrial
weather system or the Milky Way galaxy.
The historical–etiological account of function, by contrast, explicitly
restricts attribution of functions to traits that have in some sense been
designed to produce the effect cited as the function. On this account of
function, functions are the upshot of prior processes of selection. A trait has
a function if it has been designed by some process of selection to produce
some effect. In the central cases, where the traits in question are biological
 Prospects and Problems for Teleosemantics 11

adaptations, the selection process will be non-intentional natural selection.

An effect of a trait counts as its function if the trait has a certain history: in
the past possession of that trait produced the relevant effect, which in turn
had the consequence facilitating the reproduction of items with that trait.
In such cases, it is natural to adopt teleological terminology, and say that, in
the normal case, the trait exists because of an effect the trait can produce, or
in order to fulfil its function.
Take the streamlined shape shared by a number of large underwater
predators—dolphins, blue marlin, great white shark. Here we have a trait
(‘streamlinedness’) that serves these predators well, and they have the prop-
erty, it seems, because it serves them so well (cf. Griffiths and Sterelny 1999:
245–6). The fact that a number of different species that don’t share a
common ancestor have this trait suggests that the present members of the
different species have the streamlined shape because, in the environment
they share, that shape was reproductively more beneficial to their ancestors
than lack of it was to their ancestors’ competitors. So, in the environment in
which the adaptation arose, the shape has the function of facilitating swift
movement, which enables more successful predation.
Put as briefly and generally as possible, the etiological account says that
functionality arises because some individuals in a group acquire novel traits
with capacities that are favourable to their ability to reproduce. Such fea-
tures are transmitted to their descendants, proliferating within the group in
the process. Those features will then have as their function the exercise of
the favourable capacity.
Perhaps the main reason why teleosemantic theorists prefer the etiolo-
gical to the systems view of functions is that it offers a strong notion of
malfunction, which is something teleosemantics needs to account for mis-
presentation. We saw earlier how non-teleological semantic theories like
causal and indicator semantics have trouble explaining mispresentation. (If
states ‘represent’ whatever causes them, then how can they ever be tokened
falsely?) Teleosemantics offered to deal with this by distinguishing those
circumstances where a representation is ‘supposed’ to be present from oth-
er circumstances which may happen to cause it. In the latter kind of case,
the representation is malfunctioning, that is, not doing what it is sup-
posed to.
Clearly, this story needs a robust account of when a trait is doing what
it is supposed to do and when it is malfunctioning. It is not clear that the
systems approach can offer such an account. On the etiological approach, a
trait is supposed to do whatever its predecessors did that gave rise to a repro-
ductive advantage, and the trait is malfunctioning when it doesn’t do this.
By contrast, all that the systems account can offer is a statistical criterion:
in most systems of a certain kind this kind of trait does F, so here the trait
12 Graham Macdonald and David Papineau

is malfunctioning in not doing F. By contrast with the etiological analys-

is, this statistical systems account seems to lack any normative content: it
doesn’t seem to show that a trait in any sense ought to be doing F; it just says
it isn’t doing F, and so is statistically unusual, but nothing more.
So the historical–etiological approach to functions has the advantage
of making the kind of principled distinction that seems necessary for a
substantial account of malfunction in general and misrepresentation in par-
ticular. However, in tying themselves to historical–etiological functions
for this reason, teleosemanticists might seem to run into a converse diffi-
culty. As we saw earlier, the systems version attributes functionality to any
traits that cooperate to produce distinctive behaviour in complex systems.
By contrast, the historical theory allows for functions only in systems that
have been subject to some process of design. When it comes to biologic-
al systems, including our human selves, the only available designer might
seem to be the process of natural selection operating intergenerationally
on gene frequencies. However, if biological functionality always derives in
this way from genetic selection, it is surely unlikely that all representation,
including such paradigms as human beliefs and desires, can be explained in
terms of functions. After all, most human beliefs and desires are products of
ontogeny rather than phylogeny, in the sense that no genes have been selec-
ted because they foster those specific beliefs or desires. So there seems no
possibility of explaining the contents of these states in terms of etiological
design-based functions.
Fortunately for the teleosemantic project, the historical account need not
restrict functionality to traits that are genetically based in the sense that spe-
cific genes have been selected because they give rise to those traits. There are
ways in which biological items can be the products of design even though
they have no specific genetic basis in this sense. In particular, there are two
theoretical resources which often go unnoticed in this context. The first
resource, utilized most by Millikan, appeals to a many-layered account of
functions. The second resource depends on the idea of non-genetic selec-
tion. These resources greatly expand the range of items which possess eti-
ological design-based functions.

Millikan Functions
Central to the etiological account is the idea that individuals gain functional
traits as a result of being replicated. Millikan (1984, 1993) offers a highly
abstract account of replication. A simplified version goes like this: item A is
a reproduction of individual B if and only if B has some determinate properties
in common with A, and this correlation of properties can be explained by
a natural law. These common properties are the reproductively established
 Prospects and Problems for Teleosemantics 13

properties of B, and the items sharing these properties form reproductively

established families (‘refs’). Genes can form such refs. Other items that are
inherited as a result of genetic replication, such as eyes or hearts, can form
higher-order refs (horefs). ‘Direct proper functions’ are etiological functions
of traits possessed by items that are members of refs or horefs, where such
functions are the result of past selection in those families.
Millikan’s term for etiological functions is ‘proper’ function. She notes
that one kind of proper function is a relational proper function, which is a
function to do something only when bearing a certain relation to something
else. Many fish, shrimp, and prawns can adjust their colour and pattern
to the environment in which they find themselves. One prawn, Hippolyte
varians, is called the ‘chameleon prawn’ on account of its ability to adapt
its colour to its current environment across a range of yellows, greens, and
browns (see Stephenson and Stewart 1955). The prawn’s camouflage mech-
anism has the relational function of making its colour match that of its
environment, whatever that colour may be. Given a specific colour to adapt
to, the mechanism then acquires an adapted proper function. So when
the prawn is sitting on a particular brown weed, say, the adapted prop-
er function of the mechanism is to make the prawn a matching shade of
brown. Crucial here is that this precise colour may have never been pro-
duced before, so it is not a member of a ‘ref’, and the production of this
particular colour is not a direct proper function of the camouflage mechan-
ism, nor of anything else.
One can extend this picture to novel representations within composi-
tional syntactical system. Consider the famous dance of the bees, which acts
as a signal to other bees, ‘telling’ them where to go to find nectar. This
dance is adapted to the location of the nectar, so it has an adapted function.
Again, the dance indicating this specific direction may never have occurred
before. Rather, it owes its functionality to the syntax of a system that has
been reproduced because possession of such a system has yielded a repro-
ductive advantage in the past. (See Millikan, Chapter 5 in this volume, for
further extensions of this approach).

Non-Genetic Selection
The first resource enabling extension of the etiological strategy can be called
derivative functionality: devices can have the (direct) function of producing
effects that themselves have (derived) functions. The second resource avail-
able to the teleosemanticist is non-genetic selection.
So far we have not paused to analyse the notion of selection. In fact this
notion applies much more broadly than to genetically based selection. All it
requires is a set of items that that have the characteristics of:
14 Graham Macdonald and David Papineau

(a) variability in the traits possessed,

(b) selection of items with certain traits,
(c) heritability of traits selected for.
Selection cannot take place if there is no initial variation: the same selec-
tion forces operating on a homogeneous population will have no discrimin-
ating effect. When there is variation, items will then be selected for having
some trait if that trait interacts with salient features of the environment
in such a way that other items without that trait are seen to ‘suffer’ some
loss by contrast. If these favoured traits are then transmitted to descend-
ants of the items initially having those traits, the proportion of items with
these traits will increase. Whenever these conditions are satisfied, even if no
genes are involved, then it can be said that any selected trait is functional, its
function being to produce those types of effects that lead to the differential
reproduction of items with the trait in question.
There are two different modes of non-genetic selection worth mention-
ing in this context. The first operates on normal organisms but involves
non-genetic intergenerational inheritance. Many traits are passed from par-
ents to children by channels other than the sexual transmission of genetic
material: these traits will include the possession of parasites, the products
of imprinting mechanisms, and the many cognitive and behavioural traits
acquired from parents via social learning. A number of biological theorists
are currently interested in the way in which such non-genetically inherited
traits can be naturally selected through the normal Darwinian process of
differential reproduction of organisms (Jablonka and Lamb 1999; West-
Eberhard 2003; Mameli 2004). The mere fact that these traits are trans-
mitted non-genetically does not stop their possession satisfying the three
conditions above, with advantageous traits consequently becoming preval-
ent for the standard reason that their possessors have more offspring and
those offspring inherit the traits.
Non-genetically inherited traits that become prevalent in this way will
have functions, namely, the effects that favoured them. It is possible,
though this is an area that has yet to be properly explored, that functions
of this kind could do much to explain the contents of sophisticated
mental representations. After all, it seems a natural enough thought that
certain non-genetically inherited ways of thinking are an advantage to
their possessors because they make them sensitive to certain features of the
environment. On the other hand, it remains an open question how many
features of human thought are in fact due to differential reproduction of
offspring resulting from such advantages.
The second mode of non-genetic selection is more familiar and perhaps
more directly relevant to the teleosemantic project. This involves not the
 Prospects and Problems for Teleosemantics 15

differential reproduction of organisms over generations, but the differential

reproduction of cognitive or behavioural items themselves during the devel-
opment of a given individual. Such ontogenetic selection takes place, for
example, when behaviour is moulded by experience during learning. In
such cases we can think of the items selected as having the function of pro-
ducing those effects in virtue of which they were favoured by the learning
mechanism.
This kind of ontogenetic selection has been termed ‘vicarious’ or ‘sec-
ondary’ selection by Donald Campbell (1974). Campbell’s thought is that
the relevant developmental mechanisms have themselves been selected for
by genetically based natural selection to be non-genetic selectors. They oper-
ate so as to be less severe selectors than death, permitting learning and other
adaptational processes to occur.
Campbell developed an explicit ‘blind-variation-and-selective-retention’
(BVSR) model of learning. There were three essential aspects to Campbell’s
BVSR model:
(a) mechanism(s) for introducing variation,
(b) consistent selection processes,
(c) mechanism(s) for preserving and/or propagating the selected variants.
As will be clear, Campbell’s requirements for selective learning corres-
pond precisely to the requirements specified earlier for selection in general.
Campbell thought that BVSR learning mechanisms are found through-
out nature. At the most basic level, for example, an organism may avoid
noxious substances when its chemoreceptors signal that the environment
is becoming lethal. Here the chemoreceptor mechanism selects the behavi-
oural responses, this mechanism itself having been selected for precisely this
task. More generally, genetic selection spawns BVSR learning processes,
which in turn can spawn higher such processes, all the way up the tree of
knowledge. Campbell saw BVSR learning processes as yielding economies
in the creation of knowledge. At their most sophisticated, such processes
underpin the ways we use language to impart knowledge, with language
itself functioning as a substitute for the individual organism’s perceptual
investigation of its environment.
As we observed above, the kind of ontogenetic selection dealt with by
Campbell’s model of learning will yield cognitive and behavioural items
with specific etiological functions, namely, those effects in virtue of which
they were selected. Here too we will have functionality without the selec-
tion of genes, and this again expands the range of processes which can be
subject to teleosemantic analysis.
It should be said that there is as yet not a great deal of detailed work
showing how teleosemantics might analyse sophisticated human modes of
16 Graham Macdonald and David Papineau

cognition by appealing to functions other than those deriving directly from

the selection of genes. True, Millikan has indicated how her notion of an
adapted proper function can be used to account for the representational
contents of elements in complex representational systems. And Dretske has
focused on selection in learning as one means by which to explain how
cognitive states can be teleosemantically targeted on specific contents. Still,
much remains to be done in applying teleosemantics to specifically human
modes of cognition.
Perhaps this is inevitable. Detailed analyses of representational powers
in terms of etiological functions must rest on an adequate empirical know-
ledge of the cognitive mechanisms involved. There is no question of identi-
fying the functions of cognitive items if we don’t know what kinds of
mechanisms process these items and how those mechanisms develop in
individuals. From this perspective, the teleosemantic project is not so much
a theory of content for sophisticated human representation, but a methodo-
logy which promises to explain content piecemeal, in the wake of empirical
discoveries about human cognitive architecture. Progress in teleosemant-
ic accounts of human representation will come only along with empiric-
al advances in cognitive science. We hope at least to have shown in this
section that teleosemantics has plenty of resources to offer this long-term
project.

4. SUMMARIES OF CHAPTERS

The authors of the chapters in this volume were invited to contribute

because of their manifest expertise in areas central to teleosemantics.
Their approaches to that programme are diverse, as are their opinions
as to whether it possesses the potential to be successful. As editors we
hoped for a collection of articles that would enhance the discussion
both of foundational matters and of specific problems within the general
programme. We have not been disappointed. We are confident that these
essays will provide a rich source of material stimulating much further
debate, and are grateful to all our authors for their contributions.
Close to the heart of teleosemantic theories is the thought that our lin-
guistic capacity has provided our species with a crucial cognitive advantage
over rival species. As Kim Sterelny notes (Chapter 1), one view of linguist-
ic competence is that it is ‘massively modular’, encapsulated so as to make
the role of experience limited to the ‘evocation’ of one alternative from a
pre-existent, fully specified set of alternative languages. Sterelny argues that
a modest modularity hypothesis is more plausible, given the assumption
that language would have evolved in a Darwinian incremental fashion; his
 Prospects and Problems for Teleosemantics 17

suggestion is that only the organizational, syntactic, aspect of language is

modular. The proposal that our semantic understanding is analogous to
perception, a kind of ‘natural telepathy’ in which beliefs are ‘transported’
from the speaker’s mind to the hearers’ minds, is resisted, but Sterelny does
think that ‘natural telepathy’ could work for concepts. Sterelny concludes
by warning against seeing the causal process involved in this transmission
of concepts as one requiring modular mechanisms. The demands made
on linguistic expression by an ever-changing environment necessitate more
flexibility than is allowed for on the ‘massively modular’ conception of lan-
guage.
Peter Godfrey-Smith (Chapter 2) locates teleosemantics within the tradi-
tion of giving naturalistic explanations of the semantic properties of mental
representations. He investigates one basic aspect of representation—the
taking of one item, X, to tell us something about another item, Y—seeing
this as one model for mental representation. He notes that philosophers
commonly assume three features of this model: the user of the represent-
ation is distinct from the representation, the representation must have a
‘target’ established in some way, and the representation must be isomorph-
ic, perhaps abstractly so, to what it represents. Godfrey-Smith looks at the
work of some cognitive scientists to see how they implement the basic mod-
el, noting their lack of concern over the foundational issues troubling philo-
sophers. Teleosemantics can be seen as a philosophical elaboration of the
basic model, with the foundational issues, particularly the ‘target’ problem,
to the fore. He concludes with a discussion of Ruth Millikan’s particular
approach to these topics.
Fred Dretske (Chapter 3) confronts an epistemological problem faced
by any teleosemanticist, or, more broadly, any externalist. For Dretske, the
content of a (mental) representational state is given by what it has the func-
tion of indicating about the world. Given the etiological account of what
it is to have a function, it looks as though we would need to know the rel-
evant history to know what we are thinking, a strongly counter-intuitive
result. Dretske’s response is to separate out carefully two components in
what we know when we know what we think. One is that we are thinking,
the second is what we are thinking. Introspection can deliver knowledge of
the second, but it may need some empirical inquiry (e.g. into history) to
deliver knowledge of the first.
The epistemology of our knowledge of intentional content is also
discussed by Frank Jackson (Chapter 4), who argues that teleological
theories appealing to selection cannot explain how it is that the folk
have justified opinions about intentional contents. His contention is that
such theories make contents consist of properties that the folk need never
have heard of, and for that reason cannot have justified beliefs about.
18 Graham Macdonald and David Papineau

This, the ‘folk epistemology objection’, has been aired before, so Jackson
replies to various responses to the objection (cf. Braddon-Mitchell and
Jackson 1997, 2002; Papineau 2001). One such response claims that
correlations between opaque selection states and transparent states could
allow for access to the opaque content via the transparent state. But such
a correlation, argues Jackson, won’t deliver justified opinion about the
opaque content unless the folk have knowledge of the correlation—and
in most cases, they won’t. Strengthening the correlation to identity fails
the same test—the epistemic properties of (folk-accessible) intentional
content differ from the epistemic properties of any selectional role states.
Interpreting a teleosemantic theory along the lines of functional role-
realizing state theories, where the teleosemantic state is the realizer, also
won’t do; the relevant content property will be the role property, the one
to which the folk have epistemic access.
Ruth Millikan (Chapter 5) addresses a concern that was raised some time
ago by Christopher Peacocke (1992) that the content assigned to any rep-
resentation by her teleosemantic theory would be essentially anti-realistic. It
would be this because content would be sensitive to, and only to, selection
processes, and selection can only operate on what is available to those pro-
cesses. In other words, no selection-transcendent content could be assigned
to any representation. But it seems as though we do understand content
that has not itself been selected for, content that is ‘useless’ from the point
of view of, say, reproductive advantage. Millikan shows how her teleose-
mantic approach can embrace the idea of ‘useless’ content, content that is
derived from selection processes (however these are understood) but is not
itself selected for. Her solution makes essential use of extended selection
processes: any process involving trial-and-error learning counts as a selec-
tion process. Further, any set of systematic mapping rules that are selected
for will contain the capacity to generate such ‘useless content’.
This theme of non-selected (but still teleosemantic) representational con-
tent is pursued by Dan Ryder (Chapter 6), who applies the general form of
Millikan’s notion of derived relational proper functions to the brain. Spe-
cifically he uses the idea of a modelling machine’s function to be the produc-
tion of models of those items that are fed into it (its inputs) and applies it to
the way in which natural kinds can be modelled in the brain. He shows how
a cell can become tuned to a source (a bird, say) that is the location of con-
stantly correlated features (feathers, a beak, etc.). Cellular networks tuned
in this way model the environment, which is what they were designed (via
natural selection) to do. Ryder shows how such cellular networks can meet a
particular challenge, that of showing how the extension of one concept can
be determined as different from that of another concept even though the two
extensions have superficially resembling members.
 Prospects and Problems for Teleosemantics 19

Mohan Matthen (Chapter 7) asks the question ‘What feature is repres-

ented by a perceptual experience?’ His answer is Millikanian: a perceptual
experience represents an object, say, as having feature F if that is the nor-
mal condition for the successful performance of a function of a consumer
of that representation. Is the consumer’s function univocal? Matthen claims
it is; a perceptual experience, provided by a detector of the system, has the
function of eliciting an epistemic response in the organism, where an epi-
stemic response can be as basic as altering the potentialities of connected
neurons. And he argues that stimuli evoking the same epistemic response
are alike just in case they are treated as similar by the effector (the consumer
of the representation). And the meaning (representational function) of a
perceptual experience is given by the coordination scheme that emerges in
the coevolution of the detector and effector systems.
The consumer-oriented version of teleosemantics, favoured by Millikan
and Matthen, is examined by Karen Neander (Chapter 8), who thinks it
overlooks a plausible constraint on the content assigned to any represent-
ational system. The constraint is that the assigned content should play a
role in the explanation of the behaviour, perception, or cognition of the
organism. The explanatory role is further refined: the content should be
suitable for use in explanations employed in mainstream cognitive science.
A negative point she makes is that some teleological theories, those that see
content being determined by that environmental feature representation of
which selectively favoured past users of the representation, fail to respect
this explanatory constraint. Neander argues that a theory of content must
link the content to the information processing that informs the content.
Regarding Pietrowski’s snorf-kimu example (see above), she notes that the
information processed is colour information, so any theory assigning con-
tent to the representation such as ‘snorf-freeness’ would decouple content
from information processing, and this would render it unsuitable for the
explanatory purposes of mainstream cognitive science. Neander looks in
detail at how toads distinguish their prey, and assesses the suggestion that
‘nutrient’ is the content playing a causal role in prey-catching behaviour.
This suggestion is dismissed on the grounds that the toads do not have a
capacity to detect nutrients; they do have the capacity to detect various fea-
tures of the stimulus, these being associated with nutrients often enough
for the discriminative capacity to be a fitness-enhancing capacity. Neander
then examines several possible objections to her preference for assigning the
‘narrow’ content, finding none of them convincing.
The theme of discriminatory capacities is further explored in ‘Representa-
tion and Unexploited Content’ by Robert Cummins, Jim Blackmon, David
Byrd, Alexa Lee, and Martin Roth (Chapter 9). Their claim is that any theory
of content suitable for representationalist theories in cognitive science must
20 Graham Macdonald and David Papineau

allow for the phenomenon of ‘unexploited content’, content that the system
containing it is unable to use. One may have a representation at one’s dispos-
al but be unable to exploit all of its representational features, perhaps because
we have not been taught how to use those features. Teleosemantics, they
argue, requires content to be truly ascribed only after the ability to exploit
is acquired—so for the teleosemanticist there can be no ‘unexploited’ con-
tent. They trace this failure to take into account unexploited content to a
tendency to conflate representation and indication. There are significant dif-
ferences between the two, but both can have unexploited content, content
that is there prior to selection—and, it is argued, the kind of content that is
needed for representationalist cognitive science must allow for some part of
it to be unexploited. The conclusion is that it is an objection to teleosemantic
theories that they cannot accommodate unexploited content.
As many of these chapters illustrate, teleosemantics has been used
primarily to account for cognitive aspects of mentality. Carolyn Price
(Chapter 10) applies the ‘High Church teleosemantic theory’ of Millikan
to the determination of the content of emotional appraisals, these being
intentional states (e.g. beliefs) triggering the occurrence of our emotions. A
particular problem is how to distinguish such appraisals from dispassionate
evaluative judgements. Price begins by listing various functions of
emotional appraisals, such as providing motivation, focusing attention on
relevant information, limiting the set of responses the subject will choose
from, triggering expressive behaviour. Given this, she raises the question
‘What kind of content does an emotional appraisal have—descriptive,
directive, or mixed?’ In the light of the variety of functions such appraisals
are called upon to perform, Price suggests that a mixed, descriptive and
directive, content is called for. Do evaluative judgements similarly have
mixed content? She suggests that they do, but that the directive content of
the two is different, in that the directive content of the emotional appraisal
will be more detailed about the response than will the directive content
of the evaluative judgement. The descriptive content of the emotional
appraisal will also be tied to avoidable threats, this restriction not being
applied to the descriptive content of evaluative judgements. Correlatively,
the temporal content of an emotional appraisal will be restricted to the
present, near past, or near future, whereas the temporal content of an
evaluative judgement need not be so restricted.

REFERENCES

B-M, D., and J, F. (1997), ‘The Teleological Theory of

Content’, Australasian Journal of Philosophy, 75: 474–89.
 Prospects and Problems for Teleosemantics 21

(2002), ‘A Pyrrhic Victory for Teleonomy’, Australasian Journal of Philosophy,

80: 372–7.
C, D. (1974), ‘Evolutionary Epistemology’, in P. Schilpp (ed.), Essays in
Honor of Karl Popper (La Salle, Ill.: Open Court).
C, R. (1975), ‘Functional Analysis’, Journal of Philosophy, 72: 741–65.
(2002), ‘Neo-Teleology’, in A. Ariew, R. Cummins, and M. Perlman (eds.),
Functions: New Essays in Philosophy of Psychology and Biology (Oxford: Oxford
University Press).
D, F. (1981), Knowledge and the Flow of Information (Cambridge, Mass.:
MIT Press).
(1988), Explaining Behaviour (Cambridge, Mass.: MIT Press).
(1995), Naturalising the Mind (Cambridge, Mass.: MIT Press).
F, J. (1990), ‘A Theory of Content’, in Fodor, A Theory of Content and Other
Essays (Cambridge, Mass.: MIT Press).
G, P., and S, K. (1999), Sex and Death (Chicago: University of
Chicago Press).
H, C. (1969), ‘Reduction: Ontological and Linguistic Facets’, in S. Morgen-
besser, P. Suppes, and M. White (eds.), Essays in Honor of Ernest Nagel (New
York: St Martin’s Press).
J, E., and L, M. (1999), Epigenetic Inheritance and Evolution (Oxford:
Oxford University Press).
J, F. (1998), From Metaphysics to Ethics (Oxford: Oxford University Press).
ML, B. (ed.) (1991), Dretske and his Critics (Cambridge, Mass.: Black-
well).
M, M. (2004), ‘Nongenetic Selection and Nongenetic Inheritance’, British
Journal for the Philosophy of Science, 55: 35–71.
M, R. (1984), Language, Thought, and Other Biological Categories (Cam-
bridge, Mass.: MIT Press).
(1993), White Queen Psychology and Other Essays for Alice (Cambridge, Mass.:
Bradford Books, MIT Press).
(2002), ‘Biofunctions: Two Paradigms’, in A. Ariew, R. Cummins, and
M. Perlman (eds.), Functions: New Essays in Philosophy of Psychology and Biology
(Oxford: Oxford University Press).
N, K. (1995), ‘Malfunctioning and Misrepresenting’, Philosophical Studies,
79: 109–41.
P, D. (1984), ‘Representation and Explanation’, Philosophy of Science, 61:
550–72.
(1993), Philosophical Naturalism (Oxford: Blackwell).
(2001), ‘The Status of Teleosemantics, or How to Stop Worrying about
Swampman’, Australasian Journal of Philosophy, 79: 279–89.
P, C. (1992), A Study of Concepts (Cambridge, Mass.: MIT Press).
P, P. (1992), ‘Intentional and Teleological Error’, Pacific Philosophical
Quarterly, 73: 267–82.
S, D. (1977), ‘Toward a Causal Theory of Linguistic Representation’, in
P. A. French, T. E. Uehling, Jr., and H. K. Wettstein (eds.), Midwest Studies in
22 Graham Macdonald and David Papineau

Philosophy, ii: Studies in the Philosophy of Language (Minneapolis: University of

Minnesota Press).
S, E. M., and S, C. (1955), Animal Camouflage (London:
A. & C. Black; first pub. Harmondsworth: Penguin, 1946).
W-E, M. (2003), Developmental Plasticity and Evolution (Oxford:
Oxford University Press).
W, J. T. (1990), ‘Success Semantics’, Analysis, 50: 149–57.
(1991), ‘The Normal Rewards of Success’, Analysis, 51: 65–73.
 1
Language, Modularity, and Evolution
Kim Sterelny

1. LANGUAGE AND EVOKED CULTURE

Language is at the core of the cognitive revolution that has transformed that
discipline over the last forty years or so, and it is also the central paradigm
for the most prominent attempt to synthesize psychology and evolutionary
theory. A single and distinctively modular view of language has emerged out
of both these perspectives, one that encourages a certain idealization. Lin-
guistic competence is uniform, independent of other cognitive capacities,
and with a developmental trajectory that is largely independent of envir-
onmental input (Pinker 1994, 1997). Thus language is seen as a paradigm
of John Tooby and Leda Cosmides’ concept of ‘evoked culture’: linguist-
ic experience serves only to select a specific item from a menu of innately
available options (Tooby and Cosmides 1992). In explaining this concept,
Tooby and Cosmides appeal to the metaphor of a jukebox. The human gen-
ome pre-stores a set of options, and the different experiences provided by
different cultures select different elements out of this option set. I shall argue
to the contrary that variability between speakers, the sensitivity of linguistic
development to environmental input, and the limits of encapsulation are
not noise. They are central to the language and its evolution.
Evolutionary arguments about language face a problem. Evidentially ro-
bust theories of the evolution of language are in short supply. That is no
accident. Language is unique; no other living species communicates with
even a simple language.¹ Moreover, it leaves no direct trace in the

¹ The claim that there is a qualitative difference between human language and animal
signal systems is not controversial, though there are conflicting views on the nature of the
crucial differences: see Hauser et al. (2002), Jackendoff and Pinker (2005), and Pinker
and Jackendoff (2005) for divergent views on the crucial organizational difference
between language and other systems. Hauser and his allies think that recursive structure
is the crucial novelty; Jackendoff and Pinker think that this understates the significance
of the productive lexicon. Deacon, in contrast to both groups, argues that the crucial
24 Kim Sterelny

archaeological record: communicative abilities are evident, if at all, from

their effects on material culture, local ecology, and demography. This prob-
lem cannot be completely solved. Nonetheless, a partial solution is possible
by making only conservative assumptions about language and its evolution.
For that reason I shall make no assumptions about the specific timing of
the evolution of language, nor about the specific theories of syntax and
semantics that are true of our language or its ancestors. But I shall assume
(a) human and pre-human social life was characterized by some unstable
mix of cooperation and competition; (b) within this mix, language has
played a crucial role in facilitating coordinated, cooperative behaviour; (c)
language is a complex adaptation that has evolved in a classically Darwinian
way, i.e. from simple to more complex forms by relatively small increments.
I shall not assume, however, that this evolutionary process has been entirely
biological rather than cultural;² (d ) I shall assume that language expresses a
speaker’s thoughts and that words express concepts. In turn, concepts, and
hence words, typically stand in some reference-like relationship to items in
the speaker’s environment. These assumptions are not wholly uncontrover-
sial but they are not tendentious. Most defenders of a modular conception
of language accept them (though some have rejected (c), as Pinker and
Bloom (1990) complain).
In the next two sections of this chapter I present two arguments, each
of which puts the ‘evoked culture’ model of language under pressure. First
I show a tension between this model and the idea that language evolved
incrementally. In the following section, I discuss Millikan’s ‘natural tele-
pathy’ image of language, arguing that interpreting others is not, contrary
to her suspicion, a perception-like and hence modular cognitive capacity.
In the final section, I combine these two arguments with other features of
human evolution to argue that our environment has been too unstable for
selection to hard-wire a specification of language into human minds.

2. LANGUAGE: AN ORIGIN MYTH AND ITS

CONSEQUENCES

I agree with Pinker and other defenders of the modularity hypothesis that
language is very cognitively demanding, and the evolution of language must

difference is semantic not organizational: terms in natural language are symbols, not
natural signs (Deacon 1997).
² The biology–culture dichotomy is not a good one: (Oyama et al. 2001). But by
‘biological’ I mean that phenotypic similarity is transmitted from parent to offspring
exclusively by the flow of genes; not through any form of social learning.
 Language, Modularity, and Evolution 25

indeed traverse a cognitive bottleneck. Language cannot evolve without its

cognitive preconditions evolving with it. But let me begin my softening-up
by noting that many of these cognitive demands are not plausibly con-
strued as requiring the evolution of a specialized module. For one thing,
as Merlin Donald emphasizes, language requires a great expansion of vol-
untary control over physical behaviour. Speech is voluntary. Moreover, it
demands very fine motor control (Donald 1991). Yet great-ape vocal com-
munication seems to be at best quasi-voluntary. It is motivationally tied
to emotional arousal, and cued by specific stimuli in the ape’s perceptual
world. Moreover, language makes intensive demands on memory. The dif-
ferent parties to a conversation must remember who said what to whom.
Furthermore, language is not tied to the immediate environment of the
language user; we talk about much that is removed in space and time. So
an agent must have the ability to recall lexical items, rather than having
a lexicon whose entries are accessible only through some particular per-
ceptual stimulus. The speaking ape needs a memory upgrade. Language
also makes intensive demands on attention, for speech is a complex multi-
tasking activity. In a conversation, you must do more than recall what has
been said: you need to monitor and act on the effects on your utterances and
those of others. You need to be alert to signs in the audience of a failure to
understand, loss of interest, dissent, and other clues that the conversation is
going wrong. As well, you may need to monitor the non-linguistic aspects
of the situation. For often the point of talking is to facilitate coordinated
action of some kind. Language is integrated into other aspects of social life,
and hence talking requires that we divide our attention between different
tasks, and between different aspects of our current circumstances. Many of
the cues to which we must attend are contingent, varying within and across
culture. We cannot be hard-wired with this information. Many of the capa-
cities that make language possible are not encapsulated specializations. For
they are interface capacities: their job is to integrate linguistic action within
the envelope of the agent’s overall activities and those of the social group
of which the agent is a part. As Fodor (2001) notes, these are not plausibly
construed as tasks for a module.
A defender of the evoked culture model of language might concede all
this. The effective social use of language, she might agree, imposes extra
demands on cognitive capacities that are not specialized for language. But
the evolution of language requires as well the evolution of a language mod-
ule: an innate specification of the organization of language. In my view
it is hard to see how such a module could evolve. For consider what the
evolution of language would be like, on a modular conception. Presum-
ably language will have its origins in some simple version of proto-language
(which I shall call ‘Habilene’): a system with a relatively small number
26 Kim Sterelny

of word-like elements used in short utterances whose interpretation will

be heavily dependent on contextual cues (Jackendoff 1999). Learning and
using these tools will not depend on cognitive specializations for language,
for these will not yet have evolved. Instead it will depend on the agents’
general learning and problem solving techniques. But (to continue my ori-
gin myth) even a simple system of this kind is of great benefit to those
who could acquire and use it. As Habilene became increasingly central to
hominid social life, those who failed to acquire it, and those who acquired
it imperfectly, would be at an increasingly crippling disadvantage. Their
prospects for biological success would plummet.
These factors would set the stage (a modularity theorist might conjec-
ture) for an evolutionary transformation. A phenotypic trait that was once
the result of individual learning—a trait that was a manifestation of phen-
otypic plasticity—would gradually become developmentally entrenched,
with a specific genetic basis. For the establishment of Habilene sets up selec-
tion both to accelerate its acquisition and to decouple that acquisition from
the accidents of individual experience. If Habilene were to develop through
individual experience, and if children of taciturn speakers of Habilene were
thereby to develop an impoverished capacity to use it, there would be selec-
tion for any mutation making that acquisition process less sensitive to vari-
ations in individual experience. A true (proto-)language module will evolve
in response to the socio-cultural evolution of protolanguage and its estab-
lishment as an essential feature of human social competence.
However, once established, such a module would have a profound effect
on the further evolution of language. The modularization of language de-
couples linguistic capacity from linguistic experience. Instead, experience
simply plays a triggering or selecting role. On the Tooby–Cosmides view
of evoked culture, experience just selects a particular element from the
internally specified set of available responses. To take a parallel case, on
an evoked culture model of human sexual relations, we have (in that par-
ticular jukebox) a pre-specified set of marriage and childcare customs. An
agent’s specific social experience would select the local variant from this
range of options. Likewise, in the case of language, linguistic experience sets
the parameters from a pre-specified set. In contrast, a socio-cultural model
of evolution requires that children acquiring local culture are sensitive to
variations in experience. For it is through such sensitivity that a successful
innovation in parental behaviour is transmitted to their offspring. If chil-
dren are only crudely responsive to experiences provided by their parents,
intergenerational transmission will be of low fidelity, and parental innova-
tion is unlikely to be copied to the next generation (Boyd and Richerson
1985, 1996, 2000; Richerson and Boyd 1998; Tomasello 1999a,b; Sterelny
2006).
 Language, Modularity, and Evolution 27

Thus, if complex adaptations are assembled by socio-cultural evolution-

ary mechanisms, generation N + 1 must be sensitive to the experience pro-
vided by generation N, for it is only through such sensitivity that small
improvements in the N generation can be transmitted to N + 1. Everyone
accepts that some adaptive aspects of language are built by socio-cultural
processes that depend on a culturally mediated flow of information across
the generations. For that is how specialist vocabularies must be built. At
least in this respect, language is a collective product. The vocabulary of any
language is coined, refined, and transmitted by many of its speakers. Even
in a small-scale society where every speaker ends up with full individual
command over each item of the language, no one invents their own vocab-
ulary. Language is a collective product, but one which is transmitted to
new members of the group with high fidelity; and it thus meets the condi-
tions for an evolutionary ratchet effect. Useful terminology is transmitted
accurately enough for it to be used as a basis for further improvement.³
Once an aspect of language is modularized, though, the language of the
N + 1 generation ceases to be sensitive to variations in the language use of
the N generation. Thus the modularization of language would change the
evolutionary mechanisms of language as well as the developmental mech-
anisms of language. Once a module evolves that substantially decouples
linguistic competence from variations in experience, further evolutionary
change in language can take place only through genetically based alterations
in the module which are then selected to the extent that they confer on their
bearers a selective advantage. How likely are such mechanisms to take us
from relatively simple protolanguages to full human language? In my view:
not likely at all. A module with its parameters set specifies a language, so
the members of a community amongst whom a particular module is fixed
will all speak essentially identical languages. Linguistic competence based
on a common module is uniform and insensitive to new experience. Recall,
too, that a central function of language is co-ordination. How then could
speakers of expanded varieties of language invade when they are rare? How
could, say, the ability to deploy a modal vocabulary be advantageous in
any environment in which hardly anyone else can use such notions? Those
who think language depends on an innate language module suppose that
that innate system determines the general features of our grammatical, mor-
phological, and phonological systems. If so, and if there are two different

³ Tomasello argues that this is true not just of the vocabulary of a language but
also of its structural features: many morphological features—for example, tense–aspect
systems—begin as specific items of vocabulary which are then incorporated as markers;
for example, ‘going’ and ‘will’ have been converted into future constructions in English
(Tomasello 1999; 2003: 42–4).
28 Kim Sterelny

versions of the module within the one population (Chomskysoft TM1.2 and
Chomskysoft TM1.3 ), as there must be as it evolves, then some members of the
population will speak a language that is partially incommensurable to other
members of the same group. Without the right module, even if my con-
versational partner deploys modal constructions, I will neither understand
nor come to understand them.⁴ Remember, too, that structural innova-
tions are not merely additive. If, for example, Chomskysoft TM1.3 differs from
its predecessor by having a regularized tense-aspect system modifying the
verb phrase, then sentence formation using that module will be pervasively
different from those of the predecessor. Incommensurability will not be
an occasional glitch in cross-module interactions. It will be a typical and
persistent feature of these interactions, for those equipped with the sim-
pler module cannot acquire the new structures (or, at least, cannot acquire
native-speaker ease in the use of them).
Thus, consider a case where a language module, Ancestral, is fixed in a
population. Consider a Variant that would be superior in expressive power
if it were universal in a population. It by no means follows that Variant
can invade. Even if agents are motivated to cooperate, high error rates make
coordination difficult, and the invading form may increase error rates ini-
tially, even though it would be a superior system if it were fixed in the
population. An ability to embed clauses multiply or to indicate tense and
aspect by inflections threatens just to generate extra communication fail-
ures. The fitness structure is similar to that of a small number of Tit-for-Tat
variants in an otherwise All-Defect population. Those using the Tit-for-Tat
strategy find it hard to establish, for they do worse against the majority prac-
tice than that practice does against itself. The same is true of Variant, for
the pay-off structure is as follows: Variant–Variant > Ancestral–Ancestral
> Variant–Ancestral, because Variant–Ancestral interactions will result in
more communication and coordination failures than either of the other pair-
ings. But since Variants are rare and Ancestrals are common, unless there is
sharp population segregation most Ancestral interactions will be with other
Ancestrals, and few Variant interactions will be with other Variants. Thus
Variants can invade only if (i) Variant–Variant interactions are much more
productive than Ancestral–Ancestral interactions; or (ii) Variant–Ancestral
interactions are only slightly less productive than Ancestral–Ancestral inter-
actions; or (iii) Variants rarely interact with Ancestrals.
The conditions which would allow Variant to invade may have been
satisfied in early hominid populations. But it is not likely that they were.
For there are reasons for suspecting that Variant–Ancestral interactions

⁴ Or, at best, I will understand them with more difficulty and with more misunder-
standings.
 Language, Modularity, and Evolution 29

would have had substantially worse pay-offs than Ancestral–Ancestral

interactions. For one thing, speakers of the unimproved variant may
both treat linguistic similarity and difference as a marker of group
boundaries, and be less apt to act cooperatively to those out of the group.
This is quite likely. Robin Dunbar has argued that linguistic difference
functions to mark group boundaries in this way. Dunbar suspects that
initially accidental variations between adjacent groups evolved independent
functional significance by serving as a barrier to a particular free-riding
strategy, namely, that of cheating on one’s obligations and then moving
on: the drifter’s strategy. The more it costs to shift, the less appealing that
strategy becomes. Linguistic differences, especially if they coincide with
the boundaries of trust, add to the costs of movement. In the face of
increasing linguistic difference and the increasing importance of language, a
defect–shift strategy would become increasingly unappealing⁵ (Nettle and
Dunbar 1997; Dunbar 1999). So Dunbar thinks that adjacent linguistic
communities are under selection to maintain their different identities, not
to reduce them. Whether or not the differences between languages are
adaptations to mark group differences, they very likely have that effect.
Dunbar and Nettle show that language is important to perceptions of
group identity, and that perceived group membership is important to
our dispositions to cooperate. If, then, Variant were perceived as marking
membership of a different community, then the value of Ancestral–Variant
interactions would be further degraded. For speakers of Ancestral would
be liable to think of them as ‘not one of us’ and be less motivated to act
cooperatively towards them.
There is a further factor that would add costs to Variant–Ancestral inter-
actions in this scenario. Misunderstandings impose an intrinsic cost, for
they lead to coordination failure: hence the lowered value of Ancestral–
Variant interactions even in trusting social worlds. But human social life
involves competition and manipulation as well as cooperation. Since this
fact is hardly a secret, social worlds are not always trusting. That leads to
the possibility that misunderstanding may be read as defection, and that
will increase the costs of interactions which are prone to generate misunder-
standing. Indeed, there is some reason to suspect that linguistic interactions

⁵ Moreover, accent has the further consequence that when individuals do transfer,
they carry some of their history with them. So it acts as a membership badge whether
you want one or not. Our linguistic identity is both difficult to fake and difficult to
conceal. The same is of course true of other social conventions which are both complex
and arbitrary: how to use your fish-knife; in which direction the port circulates. These
points about the role of language differences in our social life are well taken, but the
drifter strategy was probably never a genuine option for our hominid ancestors. Amongst
chimps, for example, only female adolescents have an opportunity to migrate.
30 Kim Sterelny

might be treated with particular suspicion. For defection may be especially

difficult to detect, especially in cases where other agents have information
and say nothing. Who is to know that Two-Aardvarks knew of the bear
that ate Uncle Charlie? This problem of reciprocation and policing is made
more acute by the fact that in its early stages it is likely that language use
was error prone (Deacon 1997). Before the evolution of fine motor con-
trol, speech would have been more effortful, and the sounds from which
speech was composed would have been less clearly distinguished from one
another. In earlier stages of the evolution of language, then, the recep-
tion problem was more challenging, and that challenge was met by a less
fancy system. Policing defection while maintaining a cooperative relation-
ship with genuine cooperators is much more difficult in a noisy, error-
ridden environment (Sigmund 1993). The environment, then, may well
have been somewhat more suspicious, and communication failures may
well have been more likely to erode trust, increasing the social costs of
Ancestral–Variant interactions.
There is a problem, then, if we suppose that contemporary human lan-
guage evolved in stages, with earlier populations using a simpler and less
expressively powerful version of language, and if we suppose, further, that
these earlier populations were equipped with a language module that
explained their linguistic capacities in the same way that, allegedly, a
language module explains our capacities. However, there is no invasion
problem if we suppose that these earlier populations were linguistically het-
erogeneous and developmentally plastic. Heterogeneity and plasticity both
help smooth the way for the invasion of linguistic innovators. To take
the simplest kind of example, the ability to use a 500 word vocabulary
can invade a community of speakers whose standard limit was 400, if the
speakers do not all store the same 400 lexical items. If the community stock
vocabulary is of 800 items, then the advantage of the invader would reside
in being able to communicate and coordinate on more issues with more of
the existing community. It is not (say) that the invader alone would mas-
ter modal terms; rather, the invader alone would master both modal and
functional terms. Furthermore, if the population is sensitive to linguistic
experiences, innovations can spread horizontally and obliquely and hence
users of the superior variant will more often reap the rewards of interaction
with others with the same skills. Overlapping and flexible competences in
the original community allow invasion.
Thus the invasion problem seems to rule out the following scenario. Pro-
tolanguage evolved socio-culturally, but once it became crucial for social
life it was modularized; its development became independent of variation
in linguistic experience and hence uniform. Subsequent change in fun-
damental linguistic competence took place by the selective retention of
 Language, Modularity, and Evolution 31

favourable variations in the genetic basis of language modules, not by the

socio-cultural transmission of individual innovations. This cannot be the
right picture if the evolution of new and more expressively powerful vari-
ants of language presupposes ongoing real differences in linguistic com-
petence in communities, not the invasion of new variants into uniform
communities. This still leaves open the possibility that learning and using
language has been and is mediated by language-specific adaptations. But it
does seem to undermine an evoked culture model of language—a model
in which experience acts only to select from a palette of innately specified
alternatives.
Even if I am right about the invasion problem, perhaps this does not
rule out an evoked culture model of contemporary language. Perhaps it was
not modules all the way back. On this suggestion, humans differ from our
ancestors not just in the power of our language but also in its development-
al basis. Our competence in Sapiensish depends on an innate module that
specifies its organization and expressive power. Thus all living humans speak
languages which are, specific vocabulary items aside, effectively identical
in expressive power. This, however, was an evolutionary achievement; this
cognitive uniformity was not a feature of our ancestors’ linguistic lives. The
evolution of language is the evolution of a gradually modularizing capacity.
Earlier forms of language were cruder, less expressively powerful, and more
dependent on experience for their acquisition. While I have no knock-
down objection to this suggestion, there are problems with this picture.
How could this underlying genetic and cognitive uniformity evolve in the
face of linguistic diversity and geographic dispersal? Linguistic heterogen-
eity even within a single community was not mere noise: it was essential to
language dynamics. If humans were dispersed across a number of habitats,
that would have generated further diversity, especially in non-functional
aspects of language. To put the point in a nutshell: how could the most dis-
persed hominid, the most culturally varied hominid, and the one for whom
the cultural transmission of similarity was the most important, become the
most linguistically uniform? What could homogenize heterogeneity, espe-
cially in the face of cultural mechanisms that maintain differences once
they are established. Moreover, this suggestion seems inconsistent with the
standard version of the poverty of the stimulus argument for the modular-
ity of language. For on this suggestion, full human language (or something
close to it) evolved by socio-cultural mechanisms, and thence was modular-
ized, through the selection of genetic variants that increased the pace and
robustness of acquisition. This picture clearly presupposes that it is possible
to acquire human linguistic capacities without a wired-in module. For on
this view, we had language before we had a module (though perhaps we had
other adaptations for language acquisition and use).
32 Kim Sterelny

Finally, I think there is a more plausible alternative to this Baldwinesque

view of the relationship between learned and unlearned components in the
evolution of language, Eva Jablonka’s ‘assimilate and stretch’ model (Avi-
tal and Jablonka 2000). As with language, it’s a model of the evolution
of agents that face cognitive bottlenecks because they are under selection
for the elaboration of cognitively demanding skills. In selective regimes of
this kind, evolutionary changes that make some aspect of a skill unlearned
will be favoured because they free cognitive resources. Those free resources
allow the skill to become still more complex. She had in mind such traits as
complex courtship displays: complex birdsong, bowerbird bower building.
There is a feedback loop in evolutionary systems of this kind, for inde-
pendently of the average level of complexity, the more complex displays are
preferred. Suppose that a species of bowerbird is under selection for bower
quality. For an individual male bird, the more elaborate the bower, the bet-
ter. Suppose further, and plausibly enough, that learning to build a bower
is cognitively demanding. If so, there is a cognitive bottleneck: cognitive
resources constrain the capacity to build an elaborate bower. If a male must
learn everything about bowers, only simple ones can be made. To the extent
that elements of bower building are brought under genetic control, the cog-
nitive bottleneck on such birds would be eased. Their cognitive resources
would then stretch to a more complex bower. But because there is feed-
back in the system, bower building as a whole would continue to have both
innate and learned elements, even as bowers became ever more elaborate.
In this example, the feedback loop of elaboration is driven by female
preference. But the model fits language too. Feedback there is driven by the
coevolutionary connection between language and the rest of culture: each
allows the other to become more elaborate. As human cultural and social
worlds become more elaborate, there is selection for languages of greater
precision and expressive power. As languages of greater power and accur-
acy become available, human social worlds become more elaborate. I see no
reason to suppose that this coevolutionary feedback loop has shut down.
On this picture, bringing elements of language under genetic control has
contributed to the increased complexity, power, and precision of language:
with more innate, more can be learned. On this picture the expressive
power of even contemporary language is not fixed by an innate language
module, though innate elements are crucial to that power.
In short: this view of language does not involve the wholesale rejection
of a nativist, modular conception. But I do take the invasion problem to be
grounds for some caution in accepting an evoked culture model of syntax,
a model in which the role of experience is merely to select from a menu
of pre-specified alternatives by setting parameters to their appropriate set-
tings. Even so, parsing may depend in important ways on an encapsulated,
 Language, Modularity, and Evolution 33

developmentally entrenched, and specialized adaptation. It is reasonable

to conjecture that many of the organizational features of language have
been stable over an evolutionarily significant period. If so, an evolution-
arily stable subset of the total information available to an agent may suffice
to solve the parsing problem. Hence an encapsulated system can be hard-
wired into the brain of language-using agents: nature can predict much of
the information they will need for language parsing. I do not think this is
true of interpretation: of working out what others mean.

3. LANGUAGE AS A SYSTEM OF NATURAL TELEPATHY?

Few doubt that some of the cognitive skills involved in language use are
specialized and encapsulated: those involved, for example, in the phono-
logical analysis of utterances in your native language. Likewise, few argue
that all of those cognitive skills are specialized and encapsulated. Fodor
nominated the pragmatics of language use as a paradigm of unencapsu-
lated problem solving (Fodor 1983). For example, there is no telling in
advance what information you will need to work out why a speaker says
something that is obviously and spectacularly false and who therefore can-
not be saying what they mean. So no module can be designed for this task.
In this section, I will argue that identifying the literal meaning of an agent’s
utterance is more like the case of pragmatics than that of phonological ana-
lysis.
Grice, Bennett, and others have argued for a metarepresentational pic-
ture of language. Understanding the meaning of an utterance involves
recognizing and representing the speaker’s communicative intentions
(Grice 1957; Bennett 1976). Understanding utterances essentially involves
recognizing that speakers are intentional systems. Ruth Millikan does not
doubt that we sometimes interpret others via identifying their commu-
nicative intentions, especially when participants in an exchange share no
language. But she thinks that as a general model of language use this picture
is psychologically implausible. Our system of linguistic interpretation is not
designed to recognize communicative intentions. Rather, its function is to
generate the same belief in the audience that caused the agent’s utterance.⁶
For Millikan, interpretation is much more like perception than it is like
inference. The system of meaning conventions form the channel conditions
for ‘natural telepathy’. When word tokens fulfil their proper function, a
thought in the speaker’s mind appears in the mind of the audience (Millik-
an 1984, 1998).

⁶ Or to act appropriately, if the utterance is an imperative.

34 Kim Sterelny

Millikan’s picture presupposes an extremely cooperative picture of the

use of language. For a system of communication which functioned to insert
the same belief in the minds of the audience as that expressed by a speaker
would be vulnerable to deceptive manipulation. So if human social envir-
onments included a mix of cooperative and competitive interactions, her
picture is implausible (Origgi and Sperber 2000; Sterelny 2003). How-
ever, Millikan’s metaphor of natural telepathy looks much more apt when
expressed as a claim about concepts. The channel conditions of normal lan-
guage use cause the same concepts to appear in the mind of the audience
as those that are expressed in the speaker’s utterance. Two-Aardvarks tells
Erectus Jack ‘Tiger by the pond!’ Whatever else is necessary for interpret-
ation, Erectus will not have understood Two-Aardvarks unless he under-
stands what he means by ‘tiger’. Two-Aardvarks and Erectus Jack will have
understood one another only if the concept Two-Aardvarks expresses by
‘tiger’ is the concept Erectus Jack deploys in his thoughts in response to
Two-Aardvarks’s news. When all goes well, Two-Aardvarks’s utterance of
‘tiger’ causes Erectus Jack to think of tigers.
The operation of language as a system of natural telepathy depends on
the establishment of regularized, automatized, and coordinated practices of
this kind. I have a word for tigers just so long as I have a tiger concept,
and I standardly execute my communicative intentions involving tigers
using that word. I understand your term ‘tiger’ if I typically recognize
your communicative intentions about tigers when you have them, and that
recognition is caused by the covariation between your uttering the word
‘tiger’ and your having such a communicative intention concerning tigers.
How automatic and perception-like might this causal process be? Is the
recognition of meaning mediated by modular mechanisms? Let’s compare
recognition of meaning to paradigmatically perceptual processes. Consider
such examples as: the use of stereopsis to locate objects in space; estimating
the speed of oncoming objects; locating objects in space by attending to the
different times signals reach each ear. These all involve perceptual sensitiv-
ity to recondite physical features of our environment, and computationally
complex exploitation of that sensitivity. But our perceptual response does
not require that we conceptualize these facts. Erectus Jack does not have a
concept of hue or saturation just in virtue of his colour vision system’s sens-
itivity to hue and saturation. Colour vision is informationally encapsulated,
as the persistence of illusion shows. It is autonomous, operating without
cognitive decision or supervision; thus we cannot decide to see in black
and white. These factors explain how it is that we are sensitive to hue and
saturation without having concepts for hue and saturation.
Might language decoding be the same? Erectus Jack is causally respons-
ive to features of Two-Aardvarks’s cognitive and linguistic repertoire, but
 Language, Modularity, and Evolution 35

it does not follow that he has the concept of a word, concept, or symbol.
It is true that the concept–word correspondences to which Jack is sensit-
ive are less stable than the connections between hue, saturation, and colour
judgements. For colour vision depends on constant features of the phys-
ical environment, whereas the TIGER–‘tiger’ correspondence depends on
linguistic regularities. But those regularities are stable over human gener-
ations. Moreover, as with colour vision, the process is involuntary. Once
Two-Aardvarks says ‘tiger’, Jack cannot help but hear it as that term. Thus
Ruth Millikan argues that once the system of correlations or connections
has been established (and it can be established gradually, without delib-
erate engineering or consciousness), decoding can be causally sensitive to
these connections, without agents having to have thoughts about these con-
nections. You can speak and understand without being able to talk (or
think) about speaking or understanding. On this picture, understanding
the semantics of others’ utterances is assimilated to module-like processes
(though Millikan does not use this terminology): it is automatic, unreflect-
ive; not subject to central control.
I doubt that this is right. ‘Hue’, ‘saturation’, and other terms that pick
out the components of visual cognition are not elements of folk vocabulary.
The science of human colour vision is not a refined, more subtle, more pre-
cise but recognizable version of folk thought on colour experience. Likewise
syntax and phonology are not refined versions of folk thought about lan-
guage. In contrast, ‘word’ is a folk word, as is ‘means’, ‘is a name of ’, ‘is
about’, and so forth. Much of semantics is a refined, extended, and more
precise version of folk views of language and meaning. I do not think this
is an accident. Seeing does not depend on being able to talk about seeing.
For the most part, the utility of vision does not depend on our capacity
to reflect on vision. In contrast, the utility of language is more closely tied
to our capacity to reflect on language. I think there are three reasons why
this is so. First, in most circumstances we do not see, or in other ways rep-
resent, our retinal image. That image is just a causal intermediary on the
road to perceptual belief. What another agent says cannot be a mere causal
intermediary in this sense, for that would lead us to be liable to deception,
manipulation, and swallowing whole the errors of others. Even if we do
typically accept what others tell us, it is still true that what others say is eval-
uated, and hence must be represented. Furthermore, the fact that we have
been told something is often an important datum in itself. It is a clue to
what others believe and want. Second, language learning and the division of
linguistic labour seems to require that we represent language as well as use
it. In learning language, we acquire much of our vocabulary from expres-
sions rather than instances. Ostensive learning plays some role in acquis-
ition, but many terms are acquired from representations of their targets,
36 Kim Sterelny

rather than from those targets themselves. We may not need folk semantic
concepts to talk. But how could we acquire a concept from a word without
having concepts for ‘word’, ‘about’, and so on? Early language use could
not rely on smooth, well-established, uniform, and relatively context-free
use of language. Stable word–concept correlations represent coevolution-
ary achievements, not initial starting points. Many early uses of language
would have often been more like language-contact situations using pidgins.
And pidgins, as many have pointed out, rely heavily on pragmatics. They
rely on speakers thinking about what their conversational partners meant,
asking themselves ‘What did he mean by ‘‘pushpush’’. Could he really have
meant . . . ?’
Thus, whatever the literal meaning turns out to be, the processes by
which one language user identifies that literal meaning of an utterance do
not seem to be modular; they are not like perceptual processing. They are
conceptualized and they are not encapsulated.

4. LANGUAGE IN A CHANGING WORLD

One consequence of an innate language module is that it partially decouples

linguistic competence from environmental input. No one suggests, of
course, that linguistic competence does now, or ever has, developed in the
complete absence of linguistic experience. If we are equipped with an innate
language module, development is less sensitive to noise; to idiosyncrat-
ic variations in the particular experience a child might have. If a child’s
competence depends on an innate module together with a modicum of
linguistic experience, the children of taciturn parents will not grow up lin-
guistically disabled. Developmental robustness is good design for traits
which are absolutely central to human life.
So far so good. But the presumption here is that variation in the linguist-
ic input is noise not signal. We do not want development to be too robust.
In a German-speaking community, we want competence to be sensitive to
the fact that the input is in German. In general, the idea that good design
decouples important adaptations from environment input assumes that the
environment is constant. If the environment is variable in some signific-
ant respect, the development of the adaptive complex should be variable
in response. To some extent, innate systems can generate adaptive response
to variation in the environment by wiring in conditional responses. This
is the evoked culture model of innately structured yet adaptive and vari-
able response to environmental variation. Thus if human males have, over
evolutionarily significant periods, lived both in polygamous societies and
in societies which are strictly monogamous, then selection could wire in
 Language, Modularity, and Evolution 37

conditional rules: ‘if you are rich and powerful in a polygamous society,
marry many wives; if you are rich and powerful in a monogamous envir-
onment, be sure to marry a highly nubile wife of high genetic quality.’
However, there are limits on wired-in conditional responses. The range
of variation and its significance must itself be constant. So innate mod-
ules do presuppose unchanging environments, though perhaps what does
not change is the range and significance of environmental variability. Even
if humans have been widely distributed through ecospace, innate systems
can direct adaptive responses so long as the region of space humans have
occupied has been stable over time.
However, there is no reason to think that the environment of human
language has been stable, even in the extended sense discussed above. For
one thing, the physical and biological environment of human evolution
has been increasingly unstable, as a result of climate change and of the
expansion of hominid species out of their ancestral range (Potts 1996,
1998; Calvin 2002). Moreover, and perhaps more importantly, human
language capacities have coevolved with other cognitive capacities and
with the processes of culture change. Language is not the only distinctive
human cognitive capacity: we differ from our closest living kin in having
far greater powers to represent the future (Suddendorf and Corballis
1997), causal interactions (Tomasello 2000), the mental life of other
agents (Heyes 1998), moral and other norms (Boyd and Richerson 1992;
Richerson and Boyd 1998). If, as is surely likely, these capacities arose
with language rather than preceded it, then the coevolution of language
with these other cognitive faculties would have greatly altered the expressive
demands on human language. Moreover, the coevolution of cognition
with culture has built a mechanism that results in the cumulative change
of human environments. The generation-by-generation construction of
specialist vocabularies discussed earlier is an instance of a more general
process. Humans are niche constructors: we rework our own environment;
think of shelters and clothes; the domestication of animals; the use of
tools (Odling-Smee 1994; Odling-Smee et al. 2003). Moreover, our niche
construction is cumulative: generation N + 1 inherits a changed world
from generation N and further modifies the world N + 2 will inherit. So
Michael Tomasello has argued that, in contrast to the great apes, there are
three timescales that matter in understanding human minds and human
culture. The great apes have social but not cultural lives, and hence there
are only two timescales in their cognitive histories: those of phylogeny and
ontogeny. In understanding human cognition, there is a third timescale:
that of the history of culture, as complex capacities are assembled. As
these are built, they interact with and transform individual ontogeny and
biological history (Tomasello 1999a).
38 Kim Sterelny

Crucial targets of language are socio-genetic constructions. What human

agents need to discuss has not been a stable target onto which selection
could lock, encoding a stable competence that can be built into us all
once and for all. The variability of human cultures and the coevolution
of language with other cognitive capacities are sources of instability. Over
the last few hundred thousand years, there is every reason to suppose
that human cultures have varied substantially in their need to encode
information linguistically. We need language to describe (and prescribe)
norms, other minds, causal interactions, future plans. Let me mention
a few other possibilities. Consider first social rank and hierarchy. We
do not know the size of human groups 200,000 years ago, but they
were likely to be relatively small and egalitarian foraging communities.
Contrast these social worlds with the much larger and far more stratified
social worlds that have come into existence: stratifications that are in
many ways built into language. Second, consider quantitative information.
The socio-cultural invention of numerals and mathematical notation
enables contemporary humans to express quantitative information far more
precisely and extensively than pre-numerate cultures could. Third, consider
other representational media. Language now has the resources to represent
depictive representational media: pictures, diagrams, drawings. Once
more, these public and enduring representational media are important
and probably relatively recent (i.e. perhaps 35,000 years old) cultural
inventions (Mithen 1998b, 2000). But once invented they are powerful
tools in their own right, and important targets of language. Consider finally
the non-actual. We have no information, of course, on how long humans
have been storytellers. For all we know, this might be a very ancient use of
language. But the use of language for mythological and religious purposes
may not be ancient, for there is no evidence that religious belief has ancient
origins (Mithen 1998a).
Of course, a good deal of what I have just been saying is at best fairly
plausible speculation. We have no way of assigning relative dates to, say,
the emergence of our knowledge of the future, of other minds, and of full
language. But if I am right in thinking that the expressive demands on
human languages have changed in important ways, and as a consequence
human languages have differed significantly in their expressive power—if,
for example, the kinds of vocabulary vary—then the structure of the lexicon
is not a stable target onto which selection can lock, pre-wiring agents with
the information that they need. In fundamental ways, the representational
powers of language are sensitive to experience, to local culture. Language is
not merely evoked by experience.
 Language, Modularity, and Evolution 39

REFERENCES

A, E., and J, E. (2000), Animal Traditions: Behavioural Inheritance in

Evolution (Cambridge: Cambridge University Press).
B, J. (1976), Linguistic Behaviour (Cambridge: Cambridge University
Press).
B, R., and R, P. (1985), Culture and the Evolutionary Process (Chica-
go: Chicago University Press).
(1992), ‘Punishment Allows the Evolution of Cooperation (or Any-
thing Else) in Sizable Groups’, Ethology and Sociobiology, 13: 171–95.
(1996), ‘Why Culture is Common but Cultural Evolution is Rare’,
Proceedings of the British Academy, 88: 77–93.
(2000), ‘Memes: Universal Acid or a Better Mouse Trap?’, in R. Augn-
er (ed.), Darwinizing Culture: The Status of Memetics as a Science (Oxford: Ox-
ford University Press).
C, W. (2002), A Brain for All Seasons: Human Evolution and Abrupt Climate
Change (Chicago: University of Chicago Press).
D, T. (1997), The Symbolic Species: The Co-evolution of Language and the
Brain (New York: Norton).
D, M. (1991), Origins of the Modern Mind: Three Stages in the Evolution of
Culture and Cognition (Cambridge, Mass.: Harvard University Press).
D, R. (1999), ‘Culture, Honesty and the Freerider Problem’, in C. Power,
C. Knight, and R. Dunbar (eds.), The Evolution of Culture (Edinburgh: Edin-
burgh University Press).
F, J. (1983), The Modularity of Mind (Cambridge, Mass.: MIT Press).
(2001), The Mind Doesn’t Work That Way (Cambridge, Mass.: MIT Press).
G, H. P. (1957), ‘Meaning’, Philosophical Review, 66: 377–88.
H, M., C, N., and F, T. (2002), ‘The Faculty of Language:
What is It, Who has It, and How Does it Evolve?’, Science, 298: 1569–79.
H, C. M. (1998), ‘Theory of Mind in Non-Human Primates’, Behavioral and
Brain Sciences, 21: 101–48.
J, R. (1999), ‘Possible Stages in the Evolution of the Language Capa-
city’, Trends in Cognitive Science, 3/7: 272–9.
and P, S. (2005), ‘The Nature of the Language Faculty and its Implica-
tions for Evolution of Language (Reply to Fitch, Hauser and Chomsky)’, Cogni-
tion 97: 211–25.
M, R. (1984), Language, Thought, and Other Biological Categories (Cam-
bridge, Mass.: MIT Press).
(1998), ‘Language Conventions Made Simple’, Journal of Philosophy, 94/4:
161–80.
M, S. (1998a), ‘The Supernatural Beings of Prehistory and the External
Storage of Religious Ideas’, in C. Renfrew and C. Scarre (eds.), Cognition and
Material Culture: The Archaeology of Symbolic Storage (Cambridge: McDonald
Institute).
40 Kim Sterelny

M, S. (1998b), ‘A Creative Explosion? Theory of the Mind, Language and

the Disembodied Mind of the Upper Palaeolithic’, in Mithen (ed.), Creativity in
Human Evolution and Prehistory (New York: Routledge).
(2000), ‘Mind, Brain and Material Culture: An Archaeological Perspective’,
in P. Carruthers and A. Chamberlain (eds.), Evolution and the Human Mind:
Modularity, Language and Metacognition (Cambridge: Cambridge University
Press).
N, D., and D, R. (1997), ‘Social Markers and the Evolution of Recip-
rocal Exchange’, Current Anthropology, 38: 93–9.
O-S, F. J. (1994), ‘Niche Construction, Evolution and Culture’, in
T. Ingold (ed.), Companion Encyclopedia of Anthropology (London: Routledge).
L, K., and F, M. (2003), Niche Construction: The Neglected
Process in Evolution (Princeton: Princeton University Press).
O, G., and S, D. (2000), ‘Evolution, Communication and the Proper
Function of Language’, in P. Carruthers and A. Chamberlain (eds.), Evolution
and the Human Mind: Modularity, Language and Metacognition (Cambridge:
Cambridge University Press).
O, S., G, P. E., and G, R. (eds.) (2001), Cycles of Contingency:
Developmental Systems and Evolution (Cambridge, Mass.: MIT Press).
P, S. (1994), The Language Instinct: How the Mind Creates Language (New
York: William Morrow).
(1997), How the Mind Works (New York: Norton).
and B, P. (1990), ‘Natural Language and Natural Selection’, Behavioral
and Brain Sciences, 13/4: 707–84.
and J, R. (2005), ‘The Faculty of Language: What’s Special about
It’, Cognition, 95: 201–36.
P, R. (1996), Humanity’s Descent: The Consequences of Ecological Instability
(New York: Avon).
(1998), ‘Variability Selection in Hominid Evolution’, Evolutionary Anthropo-
logy, 7/3: 81–96.
R, P., and B, R. (1998), ‘The Evolution of Human Ultrasociality’,
in I. Eibl-Eibisfeldt and F. Salter (eds.), Ideology, Warfare and Indoctrinability
(Oxford: Berghahn Books).
S, K. (1993), Games of Life: Explorations in Ecology, Evolution and Beha-
viour (London: Penguin).
S, K. (2003), Thought in a Hostile World (New York: Blackwell).
(2006), ‘The Evolution and Evaluability of Culture,’ Mind and Language 21:
137–165.
S, T., and C, M. (1997), ‘Mental Time Travel and the Evol-
ution of the Human Mind’, Genetic, Social, and General Psychology Monographs,
123: 133–67.
T, M. (1999a), The Cultural Origins of Human Cognition (Cambridge,
Mass.: Harvard University Press).
(1999b), ‘The Human Adaptation for Culture’, Annual Review of Anthropo-
logy, 28: 509–29.
 Language, Modularity, and Evolution 41

(2000), ‘Two Hypotheses about Primate Cognition’, in C. Heyes and

L. Huber (eds.), Evolution of Cognition (Cambridge, Mass.: MIT Press).
(2003), Constructing a Language: A Usage Based Theory of Language Acquisi-
tion (Cambridge, Mass.: Harvard University Press).
T, J., and C, L. (1992), ‘The Psychological Foundations of Culture’,
in J. Barkow, L. Cosmides, and J. Tooby (eds.), The Adapted Mind (Oxford:
Oxford University Press).
 2
Mental Representation, Naturalism,
and Teleosemantics
Peter Godfrey-Smith

1. INTRODUCTION

The ‘teleosemantic’ program is part of the attempt to give a naturalist-

ic explanation of the semantic properties of mental representations. The
aim is to show how the internal states of a wholly physical agent could,
as a matter of objective fact, represent the world beyond them. The most
popular approach to solving this problem has been to use concepts of phys-
ical correlation with some kinship to those employed in information the-
ory (Dretske 1981, 1988; Fodor 1987, 1990). Teleosemantics, which tries
to solve the problem using a concept of biological function, arrived in the
mid-1980s with ground-breaking works by Millikan (1984) and Papineau
(1984, 1987).¹
The decade or so from the early 1980s to the early 1990s was the heyday
of the program of giving naturalistic theories of mental representation. The
work was pervaded by a sense of optimism; here was a philosophical prob-
lem that seemed both fundamental and solvable. Its solution would be a
major contribution to cognitive science, and would also contribute to many
other parts of philosophy, especially epistemology. The work was accom-
panied by skeptics and naysayers of various kinds (Stich 1983; Dennett
1987), but in many circles optimism prevailed. On some days it seemed
that a full solution might be just around the corner.
This whole program seems to have lost momentum, at least for now.
Fodor, who once had detailed solutions to offer on a regular basis, now

¹ The basic ideas of teleosemantics can also be used to try to explain the semantic
properties of public representations (Millikan 1984). But the main focus, both in the
initial work and in more recent discussions, has been mental representation. I will say
quite a bit in this chapter about how mental and public representation are related, but I
will not discuss teleosemantic treatments of public representation itself.
 Mental Representation and Naturalism 43

seems to express only a vague hope that some form of informational seman-
tics will succeed (1998: 12). Teleosemantics seems to have a fair number
of people still working on it, with various degrees of faith, as can be seen
in this volume. Millikan’s enthusiasm about her initial proposals seems
undiminished, in contrast to Fodor. But the teleosemantic program is not
insulated from the general turn away from optimism. Sometimes an idea
loses momentum in philosophy for no good reason—because of a mixture
of internal fatigue and a shift in professional fashion, for example. It is pos-
sible that this is what happened with naturalistic theories of representation.
But I think that many people have been quietly wondering for a few years
whether the naysayers might have been right all along.
More concretely, I think there is a growing suspicion that we have been
looking for the wrong kind of theory, in some big sense. Naturalistic treat-
ments of semantic properties have somehow lost proper contact with the
phenomena, both in philosophy of mind and in parts of philosophy of lan-
guage. But this suspicion is not accompanied by any consensus on how to
rectify the problem. In this chapter, my response to this difficult situation is
to re-examine some basic issues, put together a sketch of one possible altern-
ative approach, and then work forward again with the aid of this sketch.²
So a lot of the chapter is concerned with the idea of mental representa-
tion in general, and what philosophy can contribute to our understanding
of this phenomenon. These foundational discussions take up the next two
sections. Section 4 then looks at some empirical work that makes use of
the idea of mental representation—a different empirical literature from the
ones that philosophers usually focus on. Then in Section 5 I look at teleo-
semantics from the perspective established in the preceding sections.

2. REPRESENTATIONALISM REASSESSED

According to the main stream of work in naturalistic philosophy of mind

in the 1980s, inner states of organisms like us represent the world. ‘Rep-
resentation’ here is understood as a real, fairly unified natural relation that
is picked out and understood in a very vague way by folk theory, and will
eventually be described in much more detail by cognitive science and philo-
sophy. One standard form of opposition to this picture is the ‘interpretivist’

² I have undertaken similar forays in a couple of other papers in edited collections

(Godfrey-Smith 2004, 2005). There may be some tensions across these different papers,
all of which are expressed in a cautious and exploratory way. The 2004 paper discusses
problems with the mainstream 1980s program, and possible concessions to the naysayers,
in more detail.
44 Peter Godfrey-Smith

family of positions (Dennett 1987; Davidson 1984), according to which

there are no semantic properties over and above those attributed by inter-
preters, where the role of interpreter is associated with a characteristic set of
interests and point of view.
This mild caricature of a familiar clash provides a point from which to
look for new alternatives. What we want, I suggest, is a view that says
something like this: There are indeed various kinds of connectedness and
specificity that link inner states with conditions in the external world. But
we should not look so directly to the everyday concepts of representation,
belief, meaning, and so on, in describing what these connections are. The
folk apparatus of everyday interpretation is primarily a social tool. It has
genuine descriptive and explanatory uses, but these are mixed in with other
features, and it is easy to be misled by socially tuned quirks of the apparatus,
when trying to use it to describe real relations between inner states and the
world.³
In some ways, this alternative shades into each of the more standard
options mentioned above. But it is not supposed to be just a middle road.
The idea here is that it is time to consider different possible accounts of
what kinds of application semantic descriptions might have, both in prin-
ciple and in practice, to inner states of physical systems. This chapter
explores one possibility of this type.
The main idea I will discuss is that we might see the idea of mental rep-
resentation as the application of a particular model to mental phenomena.
More precisely, we might see one kind of application of the idea of mental
representation in these terms. The model in question is a schematized ver-
sion of the pattern seen in one central kind of public representation use.
That pattern is extracted and used in an attempt to understand mental
processes. I see this attempted model-based understanding of the mind as
available to the ‘folk’, and available also to scientists and philosophers who
treat the model in more serious and rigorous ways.
This model is one ‘route’ to the semantic description of inner states. It is
probably not the only one; another, possibly distinct, route is via a concept
of computation—via the idea of physical interactions that mirror logical
relations among propositions. A third way may be via information theory
in Shannon’s (1948) sense. I will leave open whether or not one of the
‘routes’ will turn out to be primary or fundamental. Certainly there come
to be connections between them (see the end of Section 4 below). In addi-
tion, my aim here is not to offer a theory of how we acquire and use the
most basic mentalistic concepts (thought, belief, pain, etc.). My focus is
specifically on the idea of representation.

³ Some of Stich’s papers (e.g. 1982) show this rather well.

 Mental Representation and Naturalism 45

The emphasis on models in this chapter is influenced by some ideas

in philosophy of science, where the distinctive properties of model-based
understanding have been much discussed in recent years (see especially
Giere 1988). The sense of ‘model’ I use in this chapter is as follows: a model
is a hypothetical structure that is supposed to bear some relevant resemb-
lance relation to some real-world system that we are trying to understand.
The hypothetical structure may in many cases be derived from another
more familiar system, though that is not essential to the strategy.
I think that many philosophers, and possibly more scientists, might ac-
cept that in some sense the idea of mental representation involves the applic-
ation to the mind of a model derived from public symbol use. But this fact
might usually be seen as not very informative. ‘Yes, sure it’s a model; now
let me get back to what I was doing.’ In this chapter I will keep the idea in
center stage. Interestingly, Wilfrid Sellars’s famous 1956 discussion of the
‘theoretical’ nature of folk psychological concepts used a very sophisticated
conception of theorizing that gave an important role to models, in a sense of
‘model’ fairly close to mine. But subsequent developments of Sellars’s idea
have not followed suit.
So let us now look at what I will call the ‘basic representationalist model’.
This is a structure—a sort of schema or scenario—that furnishes a way of
describing agents and their use of symbols to deal with the world. Our start-
ing point here is one familiar everyday sense of the term ‘representation’,
as applied to public, external objects. In this sense, a representation is one
thing that is taken to stand for another, in a way relevant to the control of
behavior or some other decision. More specifically, I take the paradigm case
here to be that when a person decides to control their behavior towards one
domain, Y, by attending to the state of something else, X. The state of X is
‘consulted’ in working out how to behave in relation to Y. This can take the
form of a conscious behavioral strategy, and it is also the topic of a famili-
ar kind of third-person interpretation. You might decide to consult a street
map to negotiate your way around a new neighborhood. Someone looking
on at you can specify both the map and the mapped domain; they say you
are using the map as a guide to a particular territory.
This chapter will look at both this very general sense of representation
and a more specific subcategory. The general class of cases is those where
some X is consulted as a guide to behavior directed on Y. The more spe-
cific category is the class of cases where this strategy involves the use of
a resemblance relation (perhaps an abstract and limited one) between X
and Y. When we consult street maps, we usually do so because we hope
to make use of a resemblance relation between map and mapped domain.
But the idea of consulting the state of one thing as a guide to another does
not always involve a resemblance relation. (Here I mean that we need not
46 Peter Godfrey-Smith

always hope for or rely on a representation relation, not merely that we

might sometimes hope for one that is not present.) In the simplest possible
case, what is consulted as a guide to behavior could be something as simple
as the value of a single binary variable. (One if by land, two if by sea.)
So far I have talked about a familiar public phenomenon. But this way
of thinking about representation seems to lend itself readily to the case of
mental states or brain structures. In this chapter I treat this as a kind of
modeling exercise; we take a familiar pattern seen in social phenomena, and
apply it to the case of thought. The ‘we’ here includes both ordinary people
and cognitive scientists looking for a more scientific handle on mental pro-
cesses.
The view developed here recalls, in several respects, Sellars’s account of
the operation of our ordinary mentalist concepts—roughly, what philo-
sophers now call folk psychology (1956/1997). Sellars imagined that folk
psychology might first appear as a theory in which inner processes were
hypothesized to resemble outward verbal discourse. The present account
is similar to Sellars’s in form, but is not the same view or even necessarily
linked closely to it. I am supposing that public representation use furnishes
a model for inner processes, but speech itself is probably not an especially
relevant kind of public representation, in this context. In addition, the
special features of propositional attitudes and their ascriptions are not the
focus here. The best way to develop the present story in detail might be to
tie it to Sellars’s account, but that possibility will not be discussed much
below.⁴
One problem with writing about this set of ideas is confusion resulting
from the profusion of things that can be called ‘models’. When we con-
sult the state of X in order to determine our behavior towards Y, it can
be natural to say that we are using X as a model of Y. This is often a
useful way of talking about the phenomenon in question. But what I am
concerned with in this chapter is the idea of taking that familiar pattern
or situation—where one thing is consulted as a guide to another—and
using that as a model for understanding some features of thought. So I will,
purely for practical reasons, never use the term ‘model’ in this chapter for an
internal or external object (my schematic ‘X’) that is being treated as a rep-
resentation of something else; I will only use the term ‘model’ when talking
about how the public phenomenon of representation use can be used as a
source of hypotheses about inner processes.
The basic representationalist model is a very natural (in the sense of
appealing) way of thinking about some aspects of the mind. I see the model

⁴ My account here is closer to Sellars’s account of sense-impressions, given at the very

end of his paper, than it is to his basic account of thoughts.
 Mental Representation and Naturalism 47

as something that ordinary folk readily turn to in describing mental pro-

cesses. Above I used the case of consulting the value of a binary variable as
the simplest possible example of the kind of phenomenon seen in the basic
representationalist model. Once we say this, it seems obvious that the vari-
able consulted could be either internal or external to the brain, as long as
the variable’s value can be read.
For those in some intellectual traditions, however, alarm bells are now
ringing. The application of representational talk of this kind to internal
states is a trap, raising the prospect of regresses, private-language prob-
lems, and more. The representationalist adopts an innocent look: ‘Surely
you can’t object to the internalization of the value of a binary variable?
Would it help if I etched it on my teeth, rather than in my brain?’ And
once the basic point about the possible internalization of a simple represent-
ation has been accepted, it seems reasonable to conjecture that the complex
structures of the brain could contain components that function as more
elaborate maps of external things, perhaps exploiting abstract resemblance
relations to coordinate behavior with the world. Indeed, some authors are
led to formulate very strong hypotheses along these lines. Here is Robert
Cummins: ‘what makes sophisticated cognition possible is the fact that
the mind can operate on something that has the same structure as the
domain it is said to cognize’ (1994: 297–8). I will return to this claim by
Cummins below. But for now let us continue discussing the basic model
itself, as opposed to versions of the view that include a role for resemblance
relations.
My aim is to treat the basic representationalist model in a way that avoids
philosophical excesses of several kinds. It is a mistake to think that there are
no prima facie foundational problems at all with the serious application of
the model to inner states (some will be discussed in the next section). But
it is also, I suggest, a mistake to think that semantic and representational
concepts are so inextricably tied to social interpretive practices that using
the model in psychology is just a massive error.
Here is another way to think of the situation. Consider the specific case
of maps, which various psychologists and philosophers have taken seriously
as akin to internal representational states. The familiar phenomena of map
use in the public arena have both an ‘empirical skeleton’ and a rich social
embedding. Here I do not mean that there must be a way of picking out in
bare causal terms all and only the events that would normally be described
as the use of maps, and assigning the maps definite semantic properties. I
mean to absorb the possibility that this is impossible, because of the open-
endedness and context-sensitivity of interpretive practices. I just mean that
at least many cases of map use have some typical local causal features. Our
habits of interpretation of these phenomena are affected by more than the
48 Peter Godfrey-Smith

empirical skeleton, but the empirical skeleton can be used as a source of

hypotheses about how the mind works. I see this as applying to both the
case of maps, and representations more generally.
So the empirical skeleton of public representation or map use might be
made the basis for a scientific understanding of the mind—in principle it
can do this, but this may or may not be a good idea. Perhaps it is a good
idea; perhaps there are special kinds of adaptive or intelligent dealing with
the world that are only made possibly by representation use, where this phe-
nomenon is found in public contexts and also in the mind. Ruth Millikan’s
theory, for example, treats internal and external signs as merely differently
located instances of the same natural kind. Alternatively, this might all be a
bad idea. One kind of anti-representationalist holds that the only empirical-
skeletal features of representational phenomena that might be found in the
mind’s workings are trivial ones. The critic may also argue that using repres-
entationalist ideas when formulating structural hypotheses about the mind
tends to lead to subtle regresses and pseudo-explanatory traps. So when we
use the representationalist model about the mind, we get very little return
and we face persistent dangers. It might, alternatively, be a good idea in
some sub-fields and at some stages in our understanding, while being mis-
leading elsewhere.
In some readers I imagine a feeling of impatience at this point. Do we
really need yet another ‘back to square one’ exercise? Surely it is perverse
to deny, at the present time, that representationalism has been fruitful in
many areas of cognitive science; the problems to work on now in this area
are problems of detail. I sympathize with one form of this impatience—a
form that accepts that representationalism is something like a model, and
insists that the model has done well in recent years. But I would add that
it is easy to work within the representationalist model without properly
resolving some acute foundational issues. (Indeed, that is one of the things
models are good for.)

3. THREE FEATURES AND A CHALLENGE

Let us look more closely at the ‘basic representationalist model’, and also
at versions that make use of a resemblance relation. In this section I will
discuss three characteristics of the model, and will also discuss in more
detail the problem of regresses and pseudo-explanations.
The first feature of the model I will discuss perhaps looks harmless, but
I will say quite a lot about it. When we have a situation that fits the basic
 Mental Representation and Naturalism 49

representationalist model, the representation being consulted must be, in

some sense, a distinct thing from whatever is consulting it. As the model has
it, one thing is used to guide behavior towards another. If we are describing
a particular situation as an instance of this phenomenon, there must be a
way of recognizing a separation between the representation and whatever is
using it. Paradigmatically, there is also some generality or portability to the
rule being used to interpret the representation, but I will not treat that as so
important here.
So if we are applying this representationalist model to the mind, we must
have some confidence that representations can in fact be separated from their
users or readers. Much of mainstream philosophy has simply accepted this.
There is a standard way of talking in philosophy of mind that treats this as no
problem. We often posit representational states, or structures, while suppos-
ing that in some sense they can be identified as distinct parts of the system.
The availability of different ‘levels of description’ is sometimes taken to allay
any worries that might arise on this front. This tendency is not exception-
less, but a great deal of representationalist talk simply assumes a separation
between a representation and something else that deals with it. This is com-
mon in teleosemantics and especially explicit in Millikan. Millikan’s account
is focused on things called ‘intentional icons’ (which include beliefs and oth-
er mental representations) that are situated ‘midway’ between ‘producer and
consumer’ mechanisms. One must make also make a separation assumption
in order to say what Cummins said in the quotation I gave in the previous
section—that intelligence requires that the mind operate on something with
the same structure as the domain it is dealing with.
In large parts of cognitive science, standard ways of talking also assume
separation, without worrying much about it. On the ‘classical’ compu-
tationalist side of cognitive science, there is a good reason for this. One
of the distinctive things about ordinary digital computers is the fact that
there is a good separation between the data stored in memory and the
processing apparatus that makes use of the data. (You can upgrade your
memory and your processor separately.) One can talk about a computer
in a way that violates the particular location of this distinction that is laid
down by the hardware; one can talk of a virtual processor with a different
structure from the one in the hardware, for example. But in the machine
itself, there is a separation of the right kind from the point of view of the
basic representationalist model. So if the mind is being seen as similar to
an ordinary digital computer, there is no reason to worry too much about
the possibility of data structures being inextricably tied to the processing.
From the point of view of mainstream cognitive science, it is presumably
50 Peter Godfrey-Smith

important and non-accidental that we have ended up building computers

with this good separation.
In less orthodox parts of cognitive science, especially parts associated
with connectionism, situated cognition, and dynamical systems, the ques-
tion of separation is more vivid.⁵ Sometimes a questioning of separation
is seen as antithetical to representationalism; sometimes instead it is just
described as ‘distributed representation’. Connectionists quite often want
to hang onto familiar kinds of representationalist talk. I do not deny that
they can do this, but they may have to depart from the basic represent-
ationalist model to do so, and this may have consequences. Sometimes it
seems that advocates of distributed representation want to talk in two ways
at once, both inside and outside the constraints of the basic model. The sep-
aration problem also has an interesting role in neuroscientific work. Talk
of ‘inner maps’ can be very appealing when talking about various cognitive
functions in an abstract way, but it is the neuroscientist who has to deal
with the possibility that no straightforward separation may appear between
‘map’ and ‘reader’.
I turn to a second feature of the basic model. When we engage in the
familiar interpretive practice outlined earlier, saying that X’s state is being
used as a guide to Y, we assume an answer to a question about specificity.
Why is it Y that is the ‘target’ here? In the everyday cases, a person can
say that it is Y that they are using X as a guide to. In the case of maps,
for example, they can say that they are treating X as a map of Y. Map-
ping talk of this kind fits into a larger assumed semantic framework, in
which maps, rules of interpretation, and target domains can be picked out
and distinguished. Clearly a somewhat different story must be told when
using the basic representationalist model to describe internal processing.
But I take it that some way of picking out the target domain must be avail-
able.
This general type of problem has been discussed extensively by Cummins
(1996). He sees giving a theory of ‘targets’ and giving a theory of what a
representation says about a target as two distinct parts of a theory of mental
content. As far as I can tell, Cummins and I do not have exactly the same
issue in mind when we talk about the problem of targets. Targets in my sense
are bigger and vaguer than they are in his; a typical target for me will not be a
particular object but a whole region of the environment. And I do not hold
out hope for a unified naturalistic theory of how targets are determined in
all real cases. But we are thinking of similar problems, clearly.
To make the problem vivid, consider a scientific case. Suppose that there
is a structure in a rat’s hippocampus that is said to be a ‘cognitive map’.

⁵ See e.g. Ramsey et al. (1991); van Gelder (1995); Clark (1997).
 Mental Representation and Naturalism 51

(This concept will be discussed some more in the next section.) The rat is
guiding its behavior, in some specific spatial task, by using this inner struc-
ture. It seems we can say that this is a case of the rat using the state of X (the
inner structure) as a guide to Y. But, of course, all the rat is doing is receiv-
ing input of various kinds, and combining this with various pre-existing
inner states to control behavior. It does not single out X, single out Y, and
decide to use the former as a guide to the latter.
From the point of view of the scientist, there is no problem here. The
rat is situated in a particular environment—a maze, for example. If the
scientist has reason to posit inner representations, he or she can say that
the representations are being used to deal with this particular maze. The
scientist applies what I will call a ‘thin behavioral’ specification of the target.
This is fine in practice, at least in simple cases. It is also rather philo-
sophically unsatisfying. It is natural from the scientist’s point of view to say
that the rat is using X as a guide to Y, but as far as the mechanics of the
situation are concerned, the ‘as a guide to Y’ claim seems extraneous. There
will also be a lot of vagueness in thin behavioral specifications of targets.
We have a different and richer specification of the target when it is picked
out explicitly by a separate representational act. Against this, it might be
argued that worrying about a richer and sharper specification of the target
is worrying about something that is not part of the ‘empirical skeleton’ of
representation use, and hence should not detain us. I will return to this issue
below.
The third issue I will discuss in this section is not an essential part of the
basic representationalist model, but is a feature of many applications and
developments of it. It is common when talking of mental representation
in ways inspired by the kinds of considerations discussed above to posit a
resemblance relation, albeit an abstract one, between representation and tar-
get. In what I regard as well-developed versions of this idea, the target itself
is not specified by the presence of a resemblance relation; the specification
of the target is a separate matter. Rather, the idea is that given that some
internal structure X is being consulted as a guide to Y, this consultation can
only be expected to be successful or adaptive to the extent that there is a
suitable resemblance relation between the two. So the goal, in some sense,
of consulting a representation is to exploit a resemblance relation between
representation and target.
At first glance, it surely seems clear that this should be regarded as an
optional feature of the representationalist model. Some and only some pub-
lic representations work via resemblance; why should this not be true also
of internal representations? However, it is quite common in this area to
52 Peter Godfrey-Smith

use the notion of resemblance far more broadly, and see the exploitation
of resemblance relations as a general or invariable feature of mental rep-
resentation. Sometimes, it seems to me, these claims are made in a way
that uses an extremely weak concept of resemblance or similarity. In other
cases, the concept of resemblance being used is not especially diluted, and
a genuinely strong claim is being expressed. The underlying line of reason-
ing might perhaps be something like this. In the public case, the available
relations between X and Y that might be exploited are roughly the three
distinguished many years ago by C. S. Peirce: resemblance, indication, and
conventionally established relations. The last of these is off the table in the
case of mental representation. The second can be assimilated to the first,
once resemblance or isomorphism is construed in a suitably abstract way.
So the only kind of relation that really matters here is resemblance.
For this or other reasons, many discussions of mental representation
extend the language of resemblance to cover a very broad class of cases. In
Randy Gallistel’s entry for ‘Mental Representation’ in the Elsevier
Encyclopedia of the Social and Behavioral Sciences (2001) he insists that all
representations exhibit an isomorphism with the represented domain. In cor-
respondence, Gallistel confirmed that cases usually discussed by philosophers
using concepts of information or indication (thermostats, fuel gauges, etc.)
are treated by him as involving abstract isomorphisms. Millikan’s teleo-
semantic theory uses concepts of mapping and correspondence in similarly
broad ways; occasionally she explicitly says that her theory vindicates the idea
that inner representations ‘picture’ or ‘mirror’ the world (1984: 233, 314).
And earlier I quoted Cummins (1994), who claimed that the exploitation
of structural similarity is the key to all sophisticated cognition.
I do not want to deny that there are some very subtle but still reasonable
notions of resemblance that may be used here, especially those employed in
logic and mathematics. My aim is not to restrict the talk of resemblance and
mapping to cases where some very obvious notion of picturing is involved.
Yet I resist the idea that some suitably abstract resemblance or isomorphism
relation is always involved in mental representation. When X is consulted
to guide behavior towards Y, this may involve the exploitation of an ante-
cedently specifiable resemblance relation, but it may not. It can be tempting
to add here that there must be some natural relation between representation
and target that makes the representation worth consulting. And from there,
it can seem that resemblance or isomorphism is the only genuine candid-
ate. But this is not so. Once we have an intelligent brain, it can generate
and adaptively manipulate representations that do not have any simple, eas-
ily exploited relation to their targets. (Strong versions of the ‘language of
thought’ hypothesis are expressions of this possibility: Fodor 1975.) For
this reason, I see no reason to accept the Cummins hypotheses that was
 Mental Representation and Naturalism 53

quoted earlier. That hypothesis arises out of a desire for an overly simple
explanation for when and why it is worth consulting X to deal with Y.
More precisely, there are (as we often find) strong and weak ways to read
the Cummins hypothesis, with the strong way unjustified and the weak way
misleading. In strong forms, the hypothesis was criticized in the previous
paragraph. In weak forms, the notion of similarity or resemblance is exten-
ded too far, and becomes post hoc in its application. (If a representation
was successfully and systematically consulted to deal with some target, there
must have been a similarity or isomorphism present of some kind . . .)
Before leaving this topic, I should note in fairness that the Cummins
hypothesis I have focused on here was expressed in a note attached as com-
mentary (1994) to a reprinting of an earlier chapter. The same ideas were
followed up in his 1996 book, but I have chosen to focus on a formula-
tion that Cummins presented in a rather ‘unofficial’ way. Secondly, I am
aware that the representational role of abstract but not-trivial resemblance
relations, especially those with mathematical description, needs a far more
detailed treatment than I have given it here.
The final topic I will discuss in this section is a general challenge to the
usefulness of the representationalist model. I call it a challenge to the ‘use-
fulness’ of the model, but the challenge is derived from stronger arguments,
often directed against the model’s very coherence. My aim here is to modify
and moderate an older form of challenge.
I argued that the empirical skeleton of public representation use might
be used as a model for some kinds of mental processing. But might it be
possible to see, in advance, reasons why this will be a bad or misleading
model? Famous arguments due to Wittgenstein (1953) and the tradition
of work following him are relevant here. One form of argument that is
especially relevant holds that if we import the basic structure of represent-
ation use into the head, we find that the reader or interpreter part of the
mechanism has to be so smart that we have an apparent regress, or pseudo-
explanation.
A version of this challenge to representationalist explanation in cognitive
science is expressed by Warren Goldfarb (1992). He is discussing a hypo-
thesis that people with perfect pitch make use of ‘mental tuning forks’. This
concept was introduced in a newspaper discussion of a piece of neuros-
cientific work on the different neural activity of people with and without
perfect pitch. Goldfarb regards the hypothesis of mental tuning forks as
pseudo-explanatory in the extreme.
Tuning forks! Are they sounding all the time? If so, what a cacophony! How does
the subject know which fork’s pitch to pick out of the cacophony when confronted
with a tone to identify? If they are not always sounding, how does she know which
one to sound when confronted with a tone?
54 Peter Godfrey-Smith

Real tuning forks give us the means to identify pitches, but they do so because
we have the practices and abilities to use them. The internal standard is supposed
to give us the means to identify items, but without practices and abilities, for
the internal standard is also meant to operate by itself, in a self-sufficient man-
ner. (If it were not, it would be otiose: why not settle for practices and abilities
themselves? . . .) (Goldfarb 1992: 114–15)
This line of thought might also be used to express a challenge to the
Cummins hypothesis that I have discussed several times in this chapter.
Cummins wants to explain intelligence by giving the mind access to some-
thing with the same structure as its target. Call this structure S. If the
mind’s problem is dealing with things that exhibit S, how does it help to
put something with S inside the head? The mind still has to detect and
respond to S, just as it did when S was outside.
When the challenge is expressed in these strong sorts of terms, the right
reply to it is to connect the representationalist model to the basic ideas
of ‘homuncular functionalism’ (Dennett 1978; Lycan 1981). The internal
representation is not supposed to be ‘self-sufficient’, to use Goldfarb’s term.
It would need a reader or interpreter; there must be something akin to
‘practices and abilities’. But the mind’s interpreter mechanism need not
have the whole set of practices and abilities of a human agent. The inter-
preter can be much less sophisticated than this (more ‘stupid’, as the hom-
uncular functionalist literature used to say), and might operate in a way
that is only somewhat analogous to a human agent using an external repres-
entation. The representationalist holds that positing this kind of separation
between a representation-like structure with an exploitable relation to a tar-
get and a subsystem to make use of that structure is a good hypothesis about
the mind. If we put these two components together, some special cognitive
capacities become possible.
So if the challenge is expressed by saying that we can see in advance
that no explanatory progress can be made with the basic representational-
ist model, then the challenge can be defused. But the fact that we have this
in-principle answer does not mean that we will necessarily make progress
in the actual world, by using the representationalist model. It may well be
that, for reasons akin to those expressed in the traditional challenge, there
is little in fact to be gained by employing the model. This will depend on
what the mind’s structure is actually like. In order to have some explanatory
usefulness, there needs to be the right kind of interaction between a rep-
resentation and reader in the mind. Putting it in homuncular functionalist
terms, the reader needs to be smart enough for its interaction with the rep-
resentation to be reader-like, but not so smart that the model collapses into
homuncularism of the bad kind.
 Mental Representation and Naturalism 55

4. INNER MAPS IN THE COGNITIVE SCIENCES

This section will look at one family of applications of the basic repres-
entational model in psychology and other cognitive sciences. The work
discussed in this section makes use of the concept of a mental or cognit-
ive map—a representational structure with some similarity or analogy to
familiar external maps, like street maps. This is obviously not the only way
to develop and apply the basic representational model in trying to under-
stand mental processes, but it is a very natural way to do so. As I noted in
Sections 2 and 3 of this chapter, there is a way of thinking about the repres-
entationalist model that leads people to think of resemblance or isomorph-
ism as a crucial relation between internal and external states. Looking for
inner map-like structures is a way to develop this idea.
The literature on inner maps is also, as I see it, a rather pure and dir-
ect way to use the basic representationalist model to think about the mind.
The literature on inner maps in the cognitive sciences is partially separate
from the tradition that emphasizes computation, logic, and language-like
representation. The empirical work on cognitive maps in question is often
(unsurprisingly) concerned with spatial skills, usually in non-linguistic ani-
mals. So this is a somewhat simpler arena in which the role of the repres-
entational model can be investigated. In particular, we do not have to worry
about the possible effects of public language capacities on the representa-
tional powers of thought.⁶
The notion of inner maps is also interesting because it seems to be a kind
of ‘attractor’ concept, one that people come back to over and over again
and from different parts of science and philosophy. There is something very
appealing about this idea, but of course it also raises in a vivid way the pit-
falls discussed at the end of the previous section. I should also emphasize
that the discussion in this section is an initial foray into this literature; I
hope to discuss it in more detail on another occasion. Here I will also dis-
cuss scientific work rather than philosophical work (see Braddon-Mitchell
and Jackson 1996 for a relevant philosophical discussion).
In psychology, the father of the idea of inner maps is E. C. Tolman
(1948). For Tolman, the hypothesis of ‘cognitive maps’ was put forward in
response to some particular forms of intelligent behavior, studied primarily
in rats and seen especially (though not exclusively) in dealing with space.
The crucial contrast that Tolman had in mind when he developed this

⁶ For some speculations on these issues that complement the present discussion, see
Godfrey-Smith (forthcoming b).
56 Peter Godfrey-Smith

idea was with strict ‘stimulus–response’ models; the hypothesis of cognitive

maps was motivated by the inability of stimulus–response models to ac-
count for what his rats could do. In his ‘sunburst maze’ experiment, for
example, a rat first learned a highly indirect route to a food source, and was
then presented with a large range of new paths, some of which led more or
less directly to the food source. Rats chose a nearly direct path much more
often than chance would predict. Tolman’s idea was often ignored in its
mid-twentieth-century context, but has since become much more influen-
tial. There has been a revival of the idea both in comparative psychology
and also in neuroscience.⁷
Earlier I distinguished a basic sense of representation in which the state
of one thing is used to guide behavior towards another, and a richer notion
in which this guidance involves a resemblance relation. Both in philosophy
and in the sciences, we find the term ‘map’ used in a range of weaker and
stronger senses. In its weakest senses, any internal representation can be
described as an internal or cognitive map. In its strongest senses, it involves
a notion of resemblance between the map and the target domain. As far as I
can tell, Tolman and many of the workers in psychology and neuroscience
who have followed him use the term ‘cognitive map’ in a way that is inter-
mediate between the weakest and strongest senses I am distinguishing. That
is, a cognitive map is not just any mental representation—it has something
extra—but perhaps it need not work via a resemblance relation with what it
represents. The term ‘map’ is primarily invoked in connection with spatial
cognition, but is sometimes used more generally.
For Tolman and others, the hypothesis of a cognitive map is often used
to express the hypothesis of some kind of cognitive sophistication, over and
above simple associative mechanisms and (especially) stimulus–response
processes. But there seems to be a family of relevant kinds of sophistica-
tion. In a good deal of the literature that I have looked at so far, the dialectic
is something like this. The researcher will have in mind a contrast between
two (or more) classes of inner mechanisms that might be used in dealing
with a behavioral problem (usually a problem involving space). One might
be a family of comparatively simple, associative mechanisms, and the other
will be a family of more sophisticated mechanisms that apparently involve a
more flexible and intelligent use of information. The term ‘cognitive map’
is associated with the latter.⁸ But the boundary between what counts as a
simpler or deflationary explanation and what counts as an explanation in

⁷ For a survey in comparative psychology that uses the concept, see Roberts (1998).
For a review of the neuroscientific work, see Jeffrey et al. (forthcoming).
⁸ Tolman himself, after contrasting his mapping idea with the stimulus–response
model, then compared more task-specific ‘strip maps’ with more flexible ‘comprehensive
maps’. He saw this as a gradient distinction. Once we have this distinction, the contrast
 Mental Representation and Naturalism 57

terms of cognitive mapping tends, perhaps, to shift around as background

knowledge changes.
One of the main contrasts that is salient here is between what is described
as an ‘integrated’ representation of spatial structure, and the possession of
special-purpose behavioral rules that can only be applied in specific circum-
stances (Mackintosh 2002). So a good deal of the empirical work tries to
find whether animals are able to use experience to come up with solutions
to problems that they have not been explicitly trained on. Many of Tol-
man’s experiments had this character, and so have various later ones. The
ability to use novel short cuts in navigation tasks is often taken to suggest
the presence of a ‘integrated’ knowledge of spatial structure, for example.
However, it turns out that there are associationist mechanisms that do pre-
dict the use of some kinds of short cut (Deutsch 1960; Bennett 1996); the
explanatory resources of the ‘simpler’ mechanisms are richer than they were
in Tolman’s time.
In some of Kim Sterelny’s philosophical work (2003) he explores the
distinction between ‘decoupled’ representations, that can be put to use in
a variety of behavioral tasks, and those that are ‘coupled’ to some specific
task or behavior. A lot of the talk of ‘integrated’ knowledge in discussions
of cognitive mapping is gesturing towards this same distinction; integration
is largely the ability to use learned information in a flexible range of tasks
and contexts.
The most important landmark after Tolman was the work that
introduced the idea of cognitive mapping into neuroscience, O’Keefe and
Nadel’s The Hippocampus as a Cognitive Map (1978). O’Keefe and Nadel
argued for the existence of a special kind of learning system, the ‘locale’
system, which constructs and uses map-like representations. O’Keefe and
Nadel also offered a specific neurological hypothesis: the hippocampus
is the part of the brain where this happens, at least in some animals.⁹
The neurological hypothesis was supported with various studies of deficits
(especially those associated with failure at harder spatial problems) and
also with the discovery of ‘place cells’ in the rat hippocampus. These cells
fire when the animal is in a certain place, but are not dependent on the
simplest place-related stimuli; they keep firing when the rat’s angle of view
is changed or it is plunged into darkness, for example.
In much of this literature, I think it would be quite accurate to say that
the idea of cognitive mapping is used as a model, in the specific sense

between strip maps and associative mechanisms is perhaps a little problematic in

Tolman’s paper.
⁹ In humans the hippocampus seems to be associated more with general laying down
of memories than with spatial cognition in particular.
58 Peter Godfrey-Smith

developed in the earlier sections of this chapter. A lot of discussion of

the distinctive features of maps is discussion at the level of semantic prop-
erties; maps are taken to be associated with specific kinds of representational
power. The psychologist’s focus is often on something like the question of
whether we have reason to think that the organism’s brain contains some-
thing with those representational properties. Some of the questions that are
most pressing from a philosophical point of view are not much discussed. In
particular, there is not a great deal of discussion of how a physical structure
in the brain might come to have the distinctive representational properties
associated with maps.
To illustrate this, let us look at how the target and separation problems
appear within this empirical work. The target problem seems to be all but
invisible. Suppose the researcher is watching the rat deal with a water maze.
(These are pools of colored water in which a platform is hidden just below
the surface in a particular location, which the rat must find and later relo-
cate.) The rat solves the problem. The researcher posits a cognitive map,
as the water maze is a hard problem that should defeat simple associative
mechanisms. What is the target of the hypothesized map? Obviously, it is
the water maze—that is what the rat is dealing with. Once one gets inside
the mindset of the empirical worker, the idea that there is a foundation-
al problem with the specification of the target seems strange. The focus of
the empirical work is a set of contrasting hypotheses about what is going
on inside the rat’s head. Whatever is in there is obviously directed, in some
sense, on what the researcher can see is the rat’s current behavioral prob-
lem—the water maze. The idea that mapping talk might be jeopardized by
the need to give some substantive story about why that is the target is quite
odd. So the cognitive scientific literature operates with what I called earlier
a ‘thin behavioral’ specification of the target. The target of a map is just
whatever the map is in fact used to deal with.
With respect to the separation problem, we see a similar lack of con-
cern to that found in much philosophical literature. The practice in the
empirical literature is often to describe inner maps within the context of
an imagined or assumed division between map and reader mechanisms,
without treating this as a very substantive hypothesis. This might be seen
as a harmless way of talking that may involve an idealization. If so, then
from the philosophical point of view this idealization may be bearing quite
a lot of weight.
So a lot of the time, a representationalist model using the idea of a map is
introduced into discussion as a means for describing and investigating some
empirically interesting distinctions, without much concern for the found-
ational problems. But some discussions do take some of the extra steps,
and ask how a neural structure could possibly have the properties posited
 Mental Representation and Naturalism 59

in the model. One way to do this is to introduce explicitly the concept of

computation. I see this concept not as itself part of the basic representa-
tionalist model, but as something additional that can be connected to it.
John O’Keefe (of O’Keefe and Nadel 1978), for example, has in his later
work tried to put the notion of cognitive mapping ‘on a firmer computa-
tional basis’ (1997: 280). His strategy is to show how a spatially organized
array of neural activity could encode a matrix of numerical values that con-
stitute a map of the environment. This map might be ‘consulted’ by the
rat in the control of behavior via mathematical manipulation of the matrix
(multiplication, inversion, and so on). So a large part of O’Keefe’s paper is
a demonstration that these mathematical manipulations of matrices are all
operations that neural circuits could, in principle, perform.
It would be interesting (and difficult) to look closely at how O’Keefe’s
computational account relates to the issue of separation, and also to the
homuncular functionalist treatment of the role of the ‘reader’ mechanisms,
as discussed in Section 3 above.

5. WHAT HAVE WE LEARNED FROM TELEOSEMANTICS?

Suppose the discussion in the preceding sections is on the right track. What
then is the status of the large body of philosophical work on naturalistic
theories of mental representation? In particular, what, if anything, have we
learned from teleosemantics? I take these two questions in turn.
The basic representationalist model is a schematic, vague sort of struc-
ture, and also one that is not usually described in rigorously naturalistic
terms. So the following question presents itself: supposing that we formu-
lated a version of the model in purely naturalistic terms, exactly what sorts
of semantic description could be given a principled basis in the model?
When this question is asked about a very simple and stripped-down ver-
sion of the model, we know from decades of philosophical work that the
available semantic descriptions will exhibit a range of indeterminacies and
breakdowns. But there is the possibility that richer versions of the model
may support more determinate and fine-grained semantic descriptions than
stripped-down versions do. So it is possible to take the basic model and
embed it in a more elaborate and detailed scenario, where all the extra com-
ponents of the scenario are described in purely naturalistic terms. We can
do this, and then ask which additional kinds of semantic description attach
plausibly to the resulting structure. For example, we can try to embed the
basic model in a surrounding context that will make a sharp and principled
notion of misrepresentation available, or the discrimination of contents
that involve coextensive concepts.
60 Peter Godfrey-Smith

What I am describing here is a kind of philosophical model-building,

which operates by augmenting and supplementing a basic representational-
ist model that has its origins outside philosophy. Even those philosophers
who are hostile to the general framework presented in this chapter will have
to agree that something like this is what a lot of work in the last twenty years
has, in fact, involved. The naturalistic philosophical literature has spent a
lot of time describing idealized hypothetical scenarios that have reasonable
or compelling descriptions in both physicalist and semantic terms. The aim
has been to describe as minimal and empirically feasible a structure as pos-
sible, while maximizing the number of features of paradigmatic semantic
description that attach plausibly to the structure.
In particular, this is how I see teleosemantics. The teleosemantic literat-
ure takes the basic model and embeds it within a biological setting. A simple
causal structure that conforms, roughly, to the basic model is embedded
within a context involving an evolutionary history and various forces of nat-
ural selection. We take the basic model, embed it in this setting, and see
how this affects our responses with respect to the semantic description of
the structures in the model. So the basic representationalist model itself has
no particular relationship to such things as natural selection and biologic-
al function. But it is possible to make use of such biological concepts to
show how a rather elaborate semantic description of the basic model can be
grounded in naturalistic factors.
One of the intuitions that has driven teleosemantics is the idea that rich
biological concepts of function pick out a special kind of involvement rela-
tion between parts of organisms and their environments. Edging even closer
to the semantic domain, there is a kind of specificity or directedness that an
evolved structure can have towards an environmental feature that figures in
its selective history. I think this idea is basically right; there is an important
kind of natural involvement relation that is picked out by selection-based
concepts of function. But this relation is found in many cases that do not
involve representation or anything close to it. It is found in the case of
enzymes, ordinary physical traits, and all the other features of organisms
that can have selective histories. There is nothing intrinsically semantic
about this involvement relation. But this sort of involvement relation can
be added to the basic representationalist model; that is what teleosemantics
does.
In keeping with the comments at the end of Section 2, I should add
that selection-based concepts of function might also be used within other
approaches to bridging semantic and physicalist description of inner pro-
cesses, besides the basic representationalist model. But let us see how selec-
tionist ideas contribute to versions of the model that has been the main
topic of this chapter.
 Mental Representation and Naturalism 61

I will focus once again on the target problem. I claimed in earlier sections
that the basic representationalist model can be employed, and is often em-
ployed, with a ‘thin behavioral’ specification of the target. In practice, there
is no problem saying that the target of the rat’s inner map (if there is one)
is the maze with which it is dealing. This idea was accepted, while noting
that from a philosophical point of view the specification of the target here
seems somewhat vague and extraneous. In a teleosemantic version of the
representationalist model, however, the target becomes far from extraneous.
This is because of the role of the target in a feedback process that shapes the
representation-using mechanisms.
I will use Millikan’s theory to illustrate this. A central concept in her
account is that of an ‘indicative intentional icon’. A wide range of semantic-
ally evaluable phenomena turn out to involve these structures, including
bee dances, indicative natural language sentences, and human beliefs. Mil-
likan says that an indicative intentional icon is a structure that ‘stands
midway’ between producer and consumer mechanisms that can both be
characterized in terms of biological function. The consumer mechanisms
modify their activities in response to the state of the icon in a way that
only leads systematically to the performance of the consumers’ biological
functions if a particular state of the world obtains. That state is (roughly)
the content of the icon. More specifically, though, if we have a set-up of
this kind then the icon is ‘supposed to map’ onto the world in a partic-
ular way—via the application of a particular rule or (mathematical sense)
function. Given the way that the consumers will respond to the state of the
icon, if the world is in such-and-such a state then the icon is supposed (in a
biological sense) to be in a corresponding state.
What this story involves, in abstract terms, is a combination of the basic
representationalist model plus a feedback process, in which relations be-
tween actions produced and the state of the world can shape the represen-
tation-using mechanisms. We suppose that the success of actions controlled
by the consultation of an inner representation is determined by the state
of some particular part of the world, and these successes and failures have
consequences for the modification of the cognitive system. The particular
feedback process that Millikan uses is biological natural selection. Within-
generation change is handled by an elaborate story about how selection
on learning mechanisms generates teleo-functional characterization of the
products of those learning mechanisms. It is the abstract idea of a suit-
able feedback process (or what Larry Wright once called a ‘consequence
etiology’) that is most relevant here, however (see Wright 1976). If we have
a feedback process of the right kind, then the representationalist model can
be employed in such a way that the specification of the target becomes a
natural part of the story—a real part of the mechanism. The aspect of the
62 Peter Godfrey-Smith

world that the organism’s inner representation is directed on is the aspect

whose different states control feedback processes shaping relevant parts of
the cognitive system, especially the ways in which representations of that
kind are produced and consumed.
There is also a more famous way in which the biological embedding of
the basic representationalist model can be used to motivate richer semantic
descriptions. This is the problem of error or misrepresentation. Much of the
original appeal of teleosemantics was its ability to employ teleo-functional
notions of purpose in order to deal with apparently normative aspects of
semantic phenomena. In particular, the biological notion of failure to per-
form a proper function was used to attack the problem of misrepresent-
ation, which had caused a lot of trouble for information-based theories.
Millikan has occasionally said that the sole or primary contribution of the
teleo-functional part of her theory (or any theory like it) is to handle this
problem. I do not think this is right. As will be clear from the discussion
in earlier sections of this chapter, I think the target problem is significant
in its own right, and is not (as it is sometimes seen) merely an aspect of,
or perspective on, the error problem. But the misrepresentation problem is
indeed one place where the teleosemantic machinery has seemed helpful.
In the light of the preceding discussion, it is possible to take a different
perspective on the use of teleo-functional concepts to deal with error. In
general, the ‘normative’ nature of semantic concepts has been overestimated
and overemphasized in recent decades. Part of this emphasis came from the
remarkable rhetorical power of Kripke’s discussions of semantic phenom-
ena in his book on Wittgenstein (Kripke 1982). Part of it came from other
places, including the teleosemantic literature itself. Semantic phenomena
do display something a bit like normativity in its more familiar senses. But
the relation between the quasi-normative aspect of semantic phenomena
and the quasi-normative forms of description made possible via selection-
based or teleonomic concepts of function is really a kind of analogy or
mirroring, not a potential reductive relationship. The ways in which mis-
representation can motivate descriptions of what ‘ought’ to have happened
and what is ‘supposed’ to be the case derive not from the empirical skeleton
of representational phenomena, but from the elaborate network of inter-
pretive social practices that surround it. (This is part of the real message of
Kripke’s book.)
Above I argued that by embedding the representationalist model in a cer-
tain kind of biological context, Millikan could give a good answer to the
target question. The fact that this embedding is one way to settle questions
about targets does not mean it is the only way. Other ways of augment-
ing or embedding the basic model may suffice to do the same kind of
job. As has repeatedly been noted, the empirical commitments behind any
 Mental Representation and Naturalism 63

ascription of content made on Millikan’s basis are strong. A selection pro-

cess of just the right kind is needed, to ‘aim’ an inner icon at a definite
aspect of the world. As I argued in an earlier paper (1994), it is possible to
doubt the widespread applicability of this pattern of explanation without
bringing in fanciful swampman-like cases, and without casting doubts on
Millikan’s interpretation of evolutionary theory and its relation to learning.
All that is needed to raise problems is a ‘noisier’ set of processes affecting the
evolution of the cognitive mechanisms in question. And in cases where it is
available, Millikan’s solution can also diverge in interesting ways from pre-
theoretic intuitions about the target of a representation. This is the message,
as I see it, of Paul Pietroski’s argument against teleosemantics (1992).
Pietroski uses an elaborate thought-experiment to make his case, but the
argument can also be described more schematically. He describes a case
where coordination with a particular environmental variable is responsible
for the success of a state of internal wiring in a simple organism, but where
the crucial environmental variable is not linked to anything the organism
can perceive in a single causal chain, but instead is linked to something the
organism can perceive by a common cause pattern. By responding to observ-
able variable Z, the organism coordinates its dealings with Y, where Z and
Y are (roughly) products of a common cause. Applying Millikan-style prin-
ciples to determine the content of an inner state by which the organism
guides its behavior, we find that the inner state will represent the state of
Y, the variable that matters in the explanation of success and failure. But
Y may be a variable whose states no organism of this kind can perceptually
discriminate in any circumstances. Pietroski makes a good case for the view
that his case shows a real gap between intuition and the application of Mil-
likan’s principles. Millikan’s account looks for the explanatory variable in
a feedback process affecting the representational machinery, and this vari-
able can be sufficiently far from the organism’s perceptual capacities (even
in optimal circumstances) that this seems an implausible target from the
standpoint of everyday interpretation. In retrospect, it is unsurprising that
there is no perfect match between the explanatory variable and the pre-
theoretic assignment of content. What I find surprising is that it takes some
real work to come up with a problem case that shows this clearly.
I have discussed the relation between Millikan’s work and the target
problem in some detail here. I will not say much about the separation prob-
lem. It will be obvious that Millikan’s account and (so far as I can tell)
other accounts in the teleosemantic literature tend to help themselves to
an assumption about separation without worrying much about the issue.
As Millikan says, to be an intentional icon something must exist ‘mid-
way between’ producer and consumer mechanisms; at least to some extent,
these components all have separable roles. To use one of her standard lists
64 Peter Godfrey-Smith

of examples, bee dances, adrenalin flows, indicative natural language sen-

tences, and human beliefs are all seen as indicative intentional icons. The
assumption that beliefs are structures that have the kind of distinct location
between producer and consumer mechanisms that we find in the case of bee
dances and adrenalin flows is a substantial assumption.
The other topic discussed in earlier sections that might be raised here is
the role of resemblance relations. I distinguished the basic representationalist
model from the more specific version where a resemblance relation between
representation and target is exploited. Interestingly, Millikan does present
her view as one in which a mapping or resemblance relation is an essential
part of the explanatory structure; she sees her account as a vindication of the
old idea that thoughts can picture the world, albeit in an abstract way (1984:
233, 314). Other teleosemantic theories (e.g. Papineau, 1987, Neander, in
this volume) do not have this feature. I have argued in previous discussions
(1994, 1996) that Millikan’s discussions of resemblance and mapping are
somewhat misleading. It is only a very attenuated sense in which her account,
if true, would vindicate the idea that thoughts picture the world. This point
can be made compactly by noting that the general notion of mapping that
Millikan uses to describe all indicative intentional icons will apply to natural
language sentences, as well as beliefs. True descriptive sentences like ‘snow is
white’ map or picture the world, too. I have handled the notion of mapping
in this chapter in such a way that map-like representation contrasts with
linguistic, conventionally mediated representation. It is possible to describe
notions of reference, satisfaction, and truth using the language of abstract
picturing, but I do not see much point in it.
In Millikan’s account, as I argued in the papers cited above, the talk of
mapping boils down to a certain kind of requirement of systematicity. Any
representation must be related to other possible representations in system-
atic ways, and the links between that set of representations and states of the
world they represent will have related systematic features. Systematicity of
this kind may well be a very important idea. I would distinguish it, how-
ever, from what I take to be much stronger uses of the idea of resemblance
of the kind seen (for example) in the Cummins hypothesis discussed earlier
in this chapter.
I am acutely aware that the issues about mapping and resemblance raised
here need a more detailed and careful treatment. That will have to wait for
another paper.
The main thing I have done in this section is give an alternative account
of the philosophical role of teleosemantics, focusing on Millikan’s version.
But some work in teleosemantics can also be seen as expressing empiric-
al hypotheses. The idea that teleosemantics is a straightforwardly empirical
theory has been asserted explicitly in the literature (e.g. Papineau 2001).
 Mental Representation and Naturalism 65

But I have in mind a slightly different set of empirical claims, and from
the point of view of the present chapter, the more philosophical and more
empirical aspects of teleosemantics have often been mixed together in com-
plicated ways.
One empirical hypothesis has remained live throughout the chapter: does
the basic representationalist model pick out an important natural kind that
figures in the causal organization of intelligent organisms? I have described
representationalism as a ‘model’, but models can be used to represent accur-
ately the real structure of the world. Models in the present sense should not
be associated, in any general way, with instrumentalism.
To that hypothesis about the basic model we can add hypotheses about
the evolutionarily embedded versions of the model that figure in teleo-
semantics. One hypothesis is especially vivid. Might it be the case that
the only natural systems that instantiate the basic representationalist mod-
el have been shaped not just by evolution in general, but by the particular
kinds of selective histories that figure in the teleosemantic literature? Might
a particular kind of natural selection be the only feasible etiology for the kind
of structure seen in the basic representational model?
Some advocates of teleosemantics might want to say, at this point, that
this was always the whole point of the program—the orientation of the
work has always been empirical, though very abstract. To me, however, it
seems that if an empirical hypothesis of this kind is supposed to be center
stage, it has not been properly separated before now from various other
ideas—ideas of the kind discussed in the earlier parts of this section.

6. CONCLUSION

This has been a somewhat sprawling chapter, but I will give a brief sum-
mary of my main points.
We may need some new views about the kind of application that repres-
entational talk has to inner states and processes. In this chapter I tried to
develop one such view, by treating representationalism as the application of
a model. This view is not offered as an account of our most basic mentalistic
concepts, although it might be linked with an account like Sellars’s. When
taken seriously in a scientific or philosophical context, the basic represent-
ationalist model does have interesting foundational problems, but these are
not insoluble in principle. Much of the literature on ‘cognitive maps’, espe-
cially in comparative psychology, is a rather pure application of the basic
representationalist model. Teleosemantics, especially Millikan’s work, can
be seen as a philosophical elaboration of the same model. Teleosemantics
embeds a version of the basic model in a detailed biological scenario, and
66 Peter Godfrey-Smith

shows how quite elaborate semantic descriptions of elements of this mod-

el can be matched up with, or grounded in, carefully described biological
factors. More of the baggage of folk semantic talk can be given a precise
analogue in a naturalistic, biological scenario than one might expect. The
biological treatment may not be the only approach to augmenting the basic
model that yields these kinds of results, however. Teleosemantics has not
uncovered a set of hidden historical assumptions that must be made when
engaging in any psychological or neuroscientific talk of mental represent-
ation. In addition, teleosemantic ideas may possibly be applied to content
without going via the basic representationalist model (a possibility I did not
much discuss).
Teleosemantics also contains, in a rather philosophically entangled way,
empirical hypotheses about the feasible etiologies by which structures that
instantiate the basic representationalist model might arise.
I doubt that teleosemantics, or any theory like it, will deliver the direct,
reductive, puff-of-papal-smoke solution that the 1980s literature envisaged.
We probably need to look for a different kind of philosophical approach
to semantic phenomena from what we are used to. But in the meantime,
we have learned quite a lot from naturalistic semantics, including teleo-
semantics, even if we have not learned what we might have originally hoped
to learn.¹⁰

REFERENCES

B, A. (1996), ‘Do Animals Have Cognitive Maps?’, Journal of Experimental

Biology, 199: 219–24.
B-M, D., and J, F. (1996), Philosophy of Mind and Cognit-
ive Science (Oxford: Blackwell).
C, A. (1997), Being There: Putting Brain, Body and World Together Again
(Cambridge, Mass.: MIT Press).
C, R. (1994), ‘Interpretational Semantics’, in S. Stich and T. Warfield
(eds.), Mental Representation: A Reader (Oxford: Blackwell).
(1996), Representations, Targets, and Attitudes (Cambridge, Mass.: MIT
Press).
D, D. (1984), Essays on Truth and Interpretation (Oxford: Oxford Univer-
sity Press).
D, T. (1997), The Symbolic Species: The Co-evolution of Language and the
Brain (New York: Norton).

¹⁰ I am indebted to Kim Sterelny for extensive comments on an earlier draft of this

chapter. This work has been influenced by discussions with Richard Francis and Alison
Gopnik.
 Mental Representation and Naturalism 67

D, D. C. (1978), Brainstorms: Philosophical Essays on Mind and Psychology

(Cambridge, Mass.: MIT Press).
(1987), The Intentional Stance (Cambridge, Mass.: MIT Press).
D, J. (1960), The Structural Basis of Behavior (Cambridge: Cambridge Uni-
versity Press).
D, F. (1981), Knowledge and the Flow of Information (Cambridge, Mass.:
MIT Press).
(1988), Explaining Behavior (Cambridge, Mass.: MIT Press).
F, J. A. (1975), The Language of Thought (New York: Crowell).
(1987), Psychosemantics (Cambridge, Mass.: MIT Press).
(1990), A Theory of Content and Other Essays (Cambridge, Mass.: MIT Press).
(1998), Concepts: Where Cognitive Science Went Wrong (Oxford: Oxford Uni-
versity Press).
G, R. (2001), ‘Mental Representation, Psychology of ’, in P. Baltes and
N. Smelser (eds.), The International Encyclopedia of the Social and Behavioral
Sciences (New York: Elsevier).
G, R. (1988), Explaining Science: A Cognitive Approach (Chicago: Chicago
University Press).
G-S, P. (1994), ‘A Continuum of Semantic Optimism’, in S. Stich
and T. Warfield (eds.), Mental Representation: A Reader (Oxford: Blackwell).
(1996), Complexity and the Function of Mind in Nature (Cambridge: Cam-
bridge University Press).
(2004), ‘On Folk Psychology and Mental Representation’, in
P. Staines (ed.), Representation in Mind: New Approaches to Mental Representation
pp. 147–62, (New York: Elsevier).
(2005), ‘Untangling the Evolution of Mental Representation’, in A. Zilhão
(ed.), Cognition, Evolution, and Rationality: A Cognitive Science for the XXIst Cen-
tury pp. 85–102, (London: Routledge).
G, W. (1992), ‘Wittgenstein on Understanding’, in P. A. French, T. E.
Uehling, and H. K. Wettstein (eds.), Midwest Studies in Philosophy, xvii: The
Wittgenstein Legacy (Notre Dame, Ind.: University of Notre Dame Press).
J, K., A, M., H, R., and C, S. (forthcoming),
‘Studies of the Hippocampal Cognitive Map in Rats and Humans’.
K, S. (1982), Wittgenstein on Rules and Private Language (Cambridge, Mass.:
Harvard University Press).
L, W. (1981), ‘Form, Function, and Feel’, Journal of Philosophy, 78: 24–50.
M, N. (2002), ‘Do Not Ask Whether They Have a Cognitive Map, but
How They Find Their Way Around’, Psicologica, 23: 165–85.
M, R. (1984), Language, Thought, and Other Biological Categories (Cam-
bridge, Mass.: MIT Press).
O’K, J. (1997), ‘Computations the Hippocampus Might Perform’, in L.
Nadel, L. A. Cooper, P. Culicover, and R. M. Harnish (eds.), Neural Connec-
tions, Mental Computation (Cambridge, Mass.: MIT Press).
and N, L. (1978), The Hippocampus as a Cognitive Map (Oxford: Claren-
don Press).
68 Peter Godfrey-Smith

P, D. (1984), ‘Representation and Explanation’, Philosophy of Science, 51:

550–72.
(1987), Reality and Representation (Oxford: Blackwell).
(2001), ‘The Status of Teleosemantics, or How to Stop Worrying about
Swampman’, Australasian Journal of Philosophy, 79: 279–89.
P, P. (1992), ‘Intentionality and Teleological Error’, Pacific Philosophical
Quarterly, 73: 267–82.
R, W., S, S., and G, J. (1991), ‘Connectionism, Eliminativism
and the Future of Folk Psychology’, in J. Greenwood (ed.), The Future of Folk
Psychology: Intentionality and Cognitive Science (Cambridge: Cambridge Univer-
sity Press).
R, W. A. (1998), Principles of Animal Cognition (Boston: McGraw-Hill).
S, W. (1956/1997), Empiricism and the Philosophy of Mind (Cambridge,
Mass.: Harvard University Press).
S, C. (1948), ‘A Mathematical Theory of Communication’, Bell System
Technical Journal, 27: 379–423, 623–56.
S, K. (2003), Thought in a Hostile World (Oxford: Blackwell).
S, S. P. (1982), ‘On the Ascription of Content’, in A. Woodfield (ed.),
Thought and Object: Essays on Intentionality (Oxford: Oxford University Press).
(1983), From Folk Psychology to Cognitive Science: The Case Against Belief
(Cambridge, Mass.: MIT Press).
T, E. C. (1948), ‘Cognitive Maps in Rats and Men’, Psychological Review,
55: 189–208.
 G, T. (1995), ‘What Might Cognition Be, If Not Computation?’, Journ-
al of Philosophy, 92: 345–81.
W, L. (1953), Philosophical Investigations, trans. G. Anscombe (New
York: Macmillan).
W, L. (1976), Teleological Explanations (Berkeley: University of California
Press).
 3
Representation, Teleosemantics, and the
Problem of Self-Knowledge
Fred Dretske

A verbal description of Clyde represents him as having three children. A

portrait of him as a young man represents him as having had a bushy
moustache. A thermometer represents him as now having a temperature of
103 ◦ F. Verbal descriptions, portraits, and thermometers are ordinary phys-
ical objects. So thinking of the mind in representational terms is the first
step in conceiving of the mind in naturalistically congenial terms—as like
other physical objects and events: measuring instruments, words, and pic-
tures. Thoughts and experiences of Clyde, just like thermometers, words,
and pictures, are physical objects that say things about Clyde that may or
may not be true.¹
Materialists will have little trouble thinking of judgments (proposition-
al attitudes, in general) in this way, but perceptual experiences (and other
qualia-laden phenomenal states) are generally regarded as more problem-
atic. A representational account of sense experience seems to me as unavoid-
able, however, as a representational theory of belief. The properties we use
to individuate and classify experiences are, after all, like the properties we
use to individuate and classify such obvious instances of representation
as stories: they are properties not (necessarily) of the experiences (stories)
themselves, but of what these experiences (stories) are experiences (stor-
ies) of. The qualities we experience, the ones that make an experience the
experience it is (shape, color, movement, pitch, texture, orientation, hard-
ness), needn’t be qualities of anything in the person’s head who is having
the experience and, therefore (given that experiences occur in the head) not
qualities of the experiences themselves. They needn’t (at least at the time

¹ I focus throughout on mental states—primarily beliefs (the propositional attitudes)

and experiences—that can most easily be interpreted in representational terms. This
makes my job a lot easier. The task of providing a naturalistic theory of the mind is
already hard enough without asking for more trouble.
70 Fred Dretske

of the experience) be qualities of anything. How is this possible? Dual-

ists have a story about how it is possible, but what can materialists say?
Materialists should say it is possible in the same way it is possible to have
stories—‘Cinderella’, for instance—depicting pumpkins transformed into
carriages without there being anything—certainly nothing between the
covers of the book (where the story is)—resembling such magical pump-
kins. The properties and situations one is consciously aware of in having
an experience resemble the properties and situations described in stories:
they are intentional entities, properties things are represented as having,
and things don’t have to have—nothing needs to have—the properties
things are represented as having.
Measuring instruments are familiar examples of such representation-
al–intentional systems. Nothing need be going 100 mph for a (malfunc-
tioning) speedometer to represent things as going that fast. Even when the
representation is veridical, it (the representation) needn’t have the proper-
ties it says or means the automobile has. Ordinarily, of course, a speedo-
meter (located in the car whose speed it represents) has the same speed it
represents the car as having, but the police have stationary devices that can
represent a car as going 100 mph. According to a representational theory,
experiences (of movement, say) are like that. The representational vehicle,
the thing in your head that represents the object you see as moving, doesn’t
(or needn’t) have the properties (movement) it represents this object as
having. That is why looking in a person’s (or bat’s) head won’t reveal the
qualities being experienced by the person (bat) in whose head one looks.
What I experience (see) when I look in another person’s head are the rep-
resentations—electrical–chemical events in gray, soggy, brain stuff, that
represent there to be a bright orange pumpkin out there. What the person
in whose brain I look experiences is a bright orange pumpkin out there.
I know of no other theory about the nature of perceptual experience
that tells this satisfying a story about the subjective—first-person vs. third-
person—aspects of experience. If experiences of pumpkins exist inside the
person seeing the pumpkins (if they don’t, why would they vanish when
the person closes her eyes?), it neatly explains why we can’t tell what other
people’s (or animals’) experiences are like just by looking in their head, at
the experiences they are having. We can’t tell—at least not just by looking
inside—for the same reason I can’t always tell what is happening in a story
by looking in the book where the story (the representation) is. If the book
is written in Chinese, I can, by looking in the book, see the representation,
the Chinese characters, but I fail to ‘see’ (understand, know, appreciate)
their meaning or content, and it is the meaning, not the symbols that have
it, in which pumpkins and carriages exist.
 Representation and Self-Knowledge 71

Since a representational theory accounts for these otherwise puzzling

facts about experience, it successfully bridges the explanatory gap. A repres-
entational theory of experience doesn’t, I admit, solve the ‘hard’ problem
of consciousness. It bridges this explanatory gap only by opening up an
equally puzzling gap somewhere else: how do electrical and chemical events
in gray soggy brain stuff manage to represent bright orange pumpkins?
We all understand how ink marks on paper—namely, the words ‘bright
orange pumpkin’—when embedded in an appropriate descriptive context,
manage to perform the trick. We—people with certain intentions and pur-
poses—give these symbols this power by making them mean bright, orange,
and pumpkin. We could, collectively, make these words mean something
else. But no one, presumably, gave the events occurring in our brains their
meaning. If, somewhere in our brains, there are representations of orange
pumpkins, these representations, unlike words in stories, are not conven-
tional, arbitrary, representations of the objects we see, hear, feel, and think
about. Their representational power is entirely independent of our pur-
poses, intentions, desires, and beliefs. So where do brains get this marvelous
power to represent (both conceptually in judgment and non-conceptually
in experience) the things we think about and perceive? How do they come
to mean orange pumpkin out there?
Good question. It is a question to which teleosemantics is supposed to
provide an answer. Teleosemantics is a theory of meaning—and, there-
fore, a theory of representation—that explains, or purports to explain, in
naturalistically approved terms, how configurations of matter in biologic-
al entities can mean (truly or falsely as the case may be) that there is an
orange pumpkin nearby. If it is successful, then, teleosemantics provides
an explanation of how arrangements of matter inside our heads can qual-
ify as beliefs about and experiences of orange pumpkins. Teleosemantics
(I’m here describing only my own version of it—see Dretske 1988, 1995)
does this by identifying the meaning (representational content) of a neuro-
biological state (organ, mechanism, etc.) with what it has the function of
indicating about the world. A (veridical) experience of a nearby orange
pumpkin is simply a physical state of (presumably) the brain that repres-
ents its external cause (in this case a pumpkin) as orange, pumpkin-shaped,
and nearby. It represents the pumpkin this way because this state has the
function of indicating the occurrence of these properties in its external
cause. If the external cause turns out to be a green cabbage, we get mis-
representation—a misperception—of the cabbage as an orange pumpkin.
A green cabbage is the object seen, but it looks like an orange pumpkin. If
there is no external cause (the representation is triggered by drugs in the
bloodstream), the experience is hallucinatory: there is nothing out there
72 Fred Dretske

(e.g. a green cabbage) that the representation represents to be an orange

pumpkin. It nonetheless still represents there to be an orange pumpkin out
there. The experience is, subjectively, the same. The world is different.
This answers the question, it solves the problem, only if there are
naturalistically acceptable ways of interpreting the key ingredients of the
recipe: function and indication. If we identify the representational content
of a neurological structure with its indicator function, we had better be
prepared to say where these functions come from and what it is (indication)
they have the function of doing, and we should be able to say this,
moreover, without taking for granted the ideas—belief and experience (and
such related notions as intention and purpose)—we are invoking functions
to understand. Otherwise the whole enterprise is circular. I understand
where the function of measuring instruments (words, diagrams, etc.) comes
from. It comes from us—from the purposes, intentions, and beliefs of
designers, builders, and users. That is what makes mercury in a glass tube
mean that Clyde’s temperature is 103◦ (whatever his actual temperature
happens to be). If, however, our brains have indicator functions, functions
that determine what we experience, think, and intend, these functions
cannot, like the functions of instruments, words, and diagrams, come
from us.
I won’t say anything about indication. I tried to do that elsewhere (Dret-
ske 1981). I’m more interested in the idea of function, the idea that binds
teleosemanticists together. In my own case, the (informational or indicat-
or) functions in question are either (for sense experience) phylogenetic or
(for belief and other conceptual states) ontogenetic (functions a state-type
acquires as a result of a certain type of learning). But those are details,
and I am not now interested in explaining (or defending) my own the-
ory of representation. Nor am I interested in criticizing other versions of
teleosemantics. I want, instead, to circle the wagons and address a com-
mon problem, a problem that, as far as I can tell, almost all teleosemanti-
cists face.² For if one traces representational power (the capacity to mean
that something is so) to functions, as teleosemanticists do, then one has to
deploy a notion of function that carries the kind of normative or intention-
al punch that psychological affairs so obviously exhibit.³ It won’t do, for

² Teleosemanticists face a lot more problems than the one I will be discussing—see,
for example, Perlman (2002), Enç (2002), and Walsh (2002) and, of course, Fodor
(1990)—but one thing at a time.
³ I’m not sure the quality we are after is normative (see Dretske 2001 for my reasons).
That will depend, I suppose, on what one means by ‘normative’. But whatever it is, it is
something capable of grounding the difference between the true and the false, right and
wrong, correct and incorrect, valid and invalid. It is, in a word, something capable of
putting the mis- into representation and the mal- into function.
 Representation and Self-Knowledge 73

instance, to operate with a notion of function that makes a thing’s func-

tion whatever it happens to be doing—even if what it happens to be doing
helps explain the behavior of the system of which it is a part.⁴ For that
doesn’t tell us what the thing is supposed to be doing. It doesn’t tell us, there-
fore, how misrepresentation (more generally, malfunctions) can occur. It
doesn’t tell us how it is possible get things wrong, make a mistake, com-
mit a fallacy, and reason incorrectly. If we are to distinguish meaning from
truth—as we must if we are to have a theory in which misrepresentation is
possible—we need functions a thing can stop performing without losing,
functions that stay fixed (as meaning does) as the world being represen-
ted (truth) changes, functions that can survive a change in the way things
function. Only by grounding this distinction between how a thing is and
how it is supposed to be can functions provide a way of understanding how
something can continue to say that p is true when it isn’t.⁵ If one oper-
ates (as I do) with indicator (informational) functions, one needs functions
(to indicate) that are durable enough to survive a failure to indicate. We
need functions that will enable us to distinguish (Grice 1957) non-natural
meaning (the sense in which ‘It is raining’ means it is raining) from natural
meaning (the sense in which wet streets mean it is raining).
I believe that only etiological functions, functions a thing has in virtue of
its history, are up to this task. Unless an object’s performance is seen in the
light of its history, there is no way it is supposed to be, no way it can get
things wrong, no way for it to be incorrect or defective. No way to make
things malfunction (Millikan 1989; Neander 1991). To say that a mechan-
ism (or bodily organ) is defective, diseased, or malfunctioning is to say that
it is not in the condition it is supposed to be in, and this ‘supposed to be’
only makes sense if there is a standard, a norm, that is independent of what
the mechanism is, in fact, doing. If you or I deliberately make something
that way for illustrative purposes, as a prototype, or as a work of abstract art,
or if it materializes randomly that way out of cosmic dust, it isn’t defective,
diseased, or malfunctioning no matter how much it looks and/or behaves
like an object (organ, mechanism, part) that (with a different history) is
defective, diseased, or malfunctioning. The standards or norms implied in
describing a mechanism (organ, etc.) as defective come from its function
or purpose, yes, but a function or purpose it has in virtue of its history. In

⁴ Roughly, the kind of functions associated with Cummins (1975) causal-role analysis
of function.
⁵ Hardcastle’s (2002) effort to give causal-role functions a normative quality is not
much help to materialists. She traces the normativity of a thing’s causal role (what she
calls ‘pragmatic’ functions) to the explanatory purposes of scientists (see pp. 152–3).
Such normativity is obviously not available to explain, at least not in naturalistic terms,
the normativity or intentionality of the mental.
74 Fred Dretske

the case of artifacts, the history that defines the function or purpose of a
thing relates to the intentions and purposes of those who design and use
it. In the case of biological organs and mechanisms, it comes from—where
else?—the evolutionary history or individual development of a system. It
is for this reason that defective things have to have a history. If they don’t,
they aren’t defective no matter how much they fail to work the way we want
them to.⁶
Injured, healthy, strained, stretched, diseased, flawed, ill, sick, damaged,
spoiled, ruined, marred, contaminated, defiled, corrupted, infected, mal-
formed —they are all like that. Nothing can be any of these things unless
it is the result of some historical process that has defined what that thing’s
function is (or the kind of thing it is—see footnote 5) and, therefore, what
it is supposed to be doing. If a thing is marred or damaged, for instance,
it departs in some degree from a standard that defines how it, or things of
that sort, should be. And so it is in psychological affairs. Without a history
(intentions of purposeful agents provide their own kind of history) there
are no mistakes. Or misrepresentations. There are, therefore, no representa-
tions.
It is for this reason that teleosemantics is committed to a historical, an
etiological, conception of functions. That, at least, is what I believe. I’m
not, however, going to spend more time arguing for it.⁷ For present pur-
poses, I will simply assume it. That will suffice because if this assumption
is false, if teleosemantics does not need history in its theory of meaning, if,
somehow, an object that materializes randomly can, at the first moment of
its existence (before it has acquired any relevant history), have an appro-
priate function and, thereby, misrepresent the world around it, then the

⁶ It is for this reason that I do not think a general goal-contribution analysis of

function (according to which an organ’s function is its contribution to the goals of
a system (see Boorse’s 2002 excellent discussion) will suffice for teleosemantics. Not
unless it has an etiological component to it. It comes closer than a causal-role analysis in
capturing the ‘normative’ character of biological (and psychological) affairs—what makes
something defective, broken, diseased, mistaken, or wrong—but without an etiological
component, it mistakenly assigns the same functions to the parts of physical duplicates
even if the duplicates materialized randomly. I don’t think the parts of a system that
materializes randomly (however much these systems resemble purposive systems) have
functions—things they are supposed to do. Boorse’s suggestion (2002: 76) that health
(one of the alleged ‘goals of living systems’) is simply the absence of pathology (which,
in turn, is the less than normal functioning of a part) doesn’t seem to me to help. How
do we tell whether something that materializes randomly is a healthy human being or a
defective chimpanzee, a monstrously deformed chipmunk or a diseased extraterrestrial
(one who would quickly die in the habitat to which it ‘belongs’)? Unless we have a way
of saying why the creature isn’t one of these other things, and I don’t see any principled
way of saying any of it, it makes little or no sense to speak of it as having a broken leg or
diseased kidneys, or, indeed, of making a mistake.
⁷ See chapter 5 of Dretske (1995) for more argument.
 Representation and Self-Knowledge 75

problem I’m going to discuss (below) is not really a problem for teleo-
semantics. So the worst-case scenario for teleosemantics is if the assumption
is true. So, in defense of teleosemantics, I assume it true in order to show
that, even on a worst-case scenario, teleosemantics is a plausible theory of
representation.
If the function of a thing and, thereby, its representational power, is
derived from history, teleosemantics is a particularly strong form of exter-
nalism about the mind. The external facts on which mental content (and,
therefore, the mental) supervenes are not only facts external to the repres-
entation, they are facts that sometimes (in the case of biological functions,
for instance) relate to the very remote past. Not only does the mind not
supervene on the current physical state of a system, it does not supervene on
the current global state of the universe. According to teleosemantics, what
we think and experience today—indeed, the fact that we think and experi-
ence anything at all today—depends not only on what is going on in us and
around us, but on events and conditions that existed long ago and (prob-
ably) far away. A physical duplicate of a conscious being, a person (?) who
lacked the appropriate history—a history that gave its internal states the
requisite functions—would not think and experience anything at all. The
internal machinery would function—causally speaking—in the same way,
but it wouldn’t have the same (or, indeed, any) function. There would be
no representations. The duplicate would be a zombie—devoid of thought
and experience.⁸
To describe this result as a problem for teleosemantics may be too gener-
ous. Some would describe it as a reductio ad absurdum. This result not only
strains the imagination of hard-headed materialists (I have heard respec-
ted materialists describe it as preposterous), it seems to many to be at odds
with the obvious fact that we know—often enough anyway—what we are
thinking and experiencing. And even if we don’t always know exactly what
we think and experience, we certainly know that we think and experience.
We know that we are conscious beings, and conscious beings (on a repres-
entational view of the mind) are beings that have thoughts and experiences.
Descartes gives us this much: the first and most indubitable fact is that we
think. But if thinking about or experiencing orange pumpkins (or anything
else, for that matter) requires us (or the mechanisms and organs in us) to
have had a certain history, as teleosemantics (on our assumption) tells us,
then we can either know, by introspection, by the fact that we are conscious
beings, that we’ve had such a history or we need to study history in order to
find out whether we are really conscious. Either way, it is absurd.

⁸ This, of course, is what Davidson’s Swampman (a physical duplicate of D. Davidson

that materialized randomly in a swamp) was meant to illustrate.
76 Fred Dretske

This epistemological argument against externalist theories of mental rep-

resentation has spawned a substantial literature. I do not propose to review
this literature (many of the seminal articles are collected in Ludlow and
Martin 1998) because, as I see it, this objection to externalism, and to
teleosemantics in particular, rests on a false assumption. Everyone (even
externalists) assume, mistakenly, that what we know by introspection is not
only (in the case of thought)⁹ what we think, the content of our current
propositional attitude, but also that we think it, the fact that we occupy
a mental state having this proposition as its content. If this assumption is
false, if what we know by introspection is that it is pumpkins one is think-
ing (wondering, worrying, deciding) about without knowing, at least not
in the same way, that one is thinking (wondering, worrying, or deciding)
about pumpkins, then there is no threat to externalism. What the teleo-
semanticist says is constituted by external, historical, relations—the fact
that one is mentally representing pumpkins—is not the fact that introspec-
tion yields: that it is pumpkins one is mentally representing.
This sounds paradoxical, I know, so let me begin my defense of it by
talking about young children—people who think but do not know, do not
even understand, what it means to think. This won’t get me quite where I
want to go, but it’s a step in the right direction. Psychologists tell us that a
typical 3-year-old does not have a developed concept of belief. These chil-
dren have beliefs, of course, and it is easy enough to find out, by asking
them the right questions, what it is they believe, but that they believe or
think these things is beyond their comprehension. They lack the concept
of thought; they are, nonetheless, authorities about what they think. They
enjoy a special kind of access to the content of their thought. If you want
to know whether 3-year-old Suzy thinks Daddy is home or the dog is loose,
just ask her. Is Daddy home? Is the dog loose? Her answers will tell you,
quite unerringly, what it is she believes. You cannot, to be sure, ask Suzy
directly, and in just these words, what it is she believes because (we are
assuming) Suzy doesn’t yet understand what it means to believe something.
But there are indirect ways of finding out. If Suzy understands what it
means for the dog to be loose, you can find out whether she thinks the dog
is loose by asking her whether the dog is loose. She is an authority, mind
you, not on whether the dog is loose (about this topic you may know better
than she), but on whether she thinks the dog is loose. Her authority on this

⁹ Throughout I use (like Descartes) ‘thought’ as a very general category. It (for me)
includes all propositional attitudes. In wondering whether P, in hoping, fearing, or
regretting that P, and in wanting or desiring that P, one is (in this broad sense) thinking
that P. To know that you think (in this broad sense) that P, then, is to know that you
occupy a mental state with intentional content. It is not merely to know that you (for
instance) think rather than fear that P.
 Representation and Self-Knowledge 77

topics is in no way diminished by her ignorance of what it means to think

the dog is loose.
Children who think (and say) that the dog is loose must, of course,
understand what it means for the dog to be loose but they needn’t under-
stand what it means to think it. They would not—indeed, they could
not—say (or even think) that they know what they think since saying
(or thinking) this requires them to refer or pick out what they think (that
the dog is loose) as something they think and, lacking an understanding
of what it means to think, they are unable to do this. Nonetheless, we
can certainly describe what they know in this way. The child knows what
it thinks—namely, that the dog is loose—despite not knowing (not even
thinking) that it thinks this.
I will be accused of ignoring scope ambiguities. It may be
intelligible—even true—to say that Suzy knows something about the
proposition—that the dog is loose —she believes, but that is not the same
as saying that Suzy knows what she believes. That would be like saying
that Suzy knows the answer to a difficult mathematical problem—that
it is, say, 24—simply because she knows what number is written on the
board (namely, 24) when 24 happens to be the answer to the problem.
Suzy may know that the number 24 is on the board, but unless she knows
this number under the description ‘the answer to the problem’ she doesn’t
know what the answer to the problem is. Likewise, Suzy may think that
the dog is loose, and she may know this proposition—that the dog is
loose —under some description or other (perhaps as ‘what I told Mommy’),
but unless she knows it under the description ‘what I think’ she doesn’t
know what she thinks. All Suzy really knows is what she told Mommy.
What she told Mommy is what she thinks, of course, but she doesn’t
know this.¹⁰
This is a fair objection. At least it is an objection that philosophers are
apt to make. So I concede the point. It is why I said at the beginning that
consideration of children, people who lack the concept of thought, would
not get me quite where I wanted to go. It only gets me to the point of hav-
ing established that there is some sense in which children can know what
it is they think without knowing that they think it. In this sense, knowing
that it is X you think does not require you to know you think X. Children

¹⁰ This point could also be expressed by talking about the difference (in the context of
knowledge or belief attributions) between attributive vs. referential uses of descriptions.
Following Boër and Lycan’s (1986: 18) description of the difference between a referential
sense of knowing who the murderer is (in which it is not necessary to know the murderer
murdered anyone) and the attributive sense (where this is necessary), we could say that
children know what they think in the referential sense (where it is not necessary to know
they think it), not the attributive sense (where it is necessary to know this).
78 Fred Dretske

can know what they think without knowing they think it in the same sense
you can know what my brother is doing without knowing it is my brother
doing it. But this, as my critic has been quick to point out and as I am now
willing to concede, only shows that Suzy knows of what she thinks that it is
that the dog is loose (the phrase ‘of what she thinks’ kept carefully outside
the that-clause that expresses what it is Suzy really knows). It does not show
that Suzy knows that what she thinks is that the dog is loose (the phrase
‘what she thinks’ here occurring inside the scope of the knowledge attribu-
tion). So it does not, not in any relevant sense, show that Suzy knows what
she thinks—much less that she knows what she thinks without knowing
she thinks it.
So, to make the next step in my argument, let me shift to a person who,
unlike a child, possesses the relevant concepts and beliefs. I will describe
an analogous situation—knowing what someone said—and suggest that it
provides a model for knowing what one thinks.
Clyde gets a telephone call from his good friend Harold. Harold tells
him that he is going on vacation for two weeks. Clyde hears him say this
and, let us suppose, hears him say it under ideal telephonic conditions (no
static, clear articulation, etc.), the kind of conditions that would ordinarily
prompt us to say that Clyde knows what Harold said. So far, I hope, there
is nothing suspicious. Now the twist. There are several people, all practical
jokers, who, quite unknown to Clyde, enjoy telephoning Clyde and imit-
ating Harold. They are very good at it. As far as Clyde can tell, the call he
received from Harold could have been from any one of these other people.
It sometimes is one of these other people. Unaware of the past deceptions,
and, therefore, the very real present possibilities they create, Clyde not only
believes (correctly as it turns out), without doubt or hesitation, that it is
Harold he is talking to, but (incorrectly as it turns out) that he knows it is
Harold.¹¹
I have Gettierized¹² Clyde’s belief that it was Harold on the phone while
leaving intact his evidence for what it was that Harold said to him. The
question I’m interested in is this: does the fact that Clyde does not know
it was Harold who said he was going on vacation mean that he doesn’t
know what Harold said to him? If asked (‘What did Harold say?’), Clyde
will tell you, confidently and truthfully, exactly what Harold said. If asked
whether he knows—and, if so, how he knows—that this is what Harold

¹¹ I here assume that if Clyde can’t tell the difference between Harold’s voice on the
phone and the voices of several other people, any one of whom might be calling, then,
whether or not he realizes it, he doesn’t know it is Harold. He certainly can’t hear that it
is Harold.
¹² That is, I have described conditions in which Clyde has a justified true belief (that
it is Harold he is talking to) that does not constitute knowledge.
 Representation and Self-Knowledge 79

said, Clyde will tell you, once again confidently and (I submit) truthfully,
that he knows this because he heard him say it. If anyone ever knows what
another person says on the phone, Clyde, given the circumstances, surely,
knows what Harold said. Yet, Clyde doesn’t know it was Harold who said
it. Clyde thinks he knows. This, indeed, is why he so confidently reports
what he knows by referring to the caller as Harold. But the truth of the
matter is that Clyde is ignorant about who called him.
Unlike the earlier case of the child, we now have an example in which
the agent does understand the phrase (‘what Harold said’) being used to
pick out the proposition that he heard expressed on the phone. He not
only understands it, he confidently (and truly!) believes it refers to what
he heard. That is why he describes what he heard the caller say as ‘what
Harold said’. Unlike the case of the child who does not believe, does not
even understand, that ‘what I think’ (when said or thought by her) is a
correct description of the mental state whose content (namely, that the
dog is loose) she has special access to, Clyde does understand—indeed, he
truly and confidently believes, that ‘what I heard Harold say’ is a correct
description of the content he has special (auditory) access to, the proposi-
tion he heard expressed on the telephone. Why isn’t this enough to know
not (once again) that it was Harold who said he was going on vacation, but
that what Harold said was that he was going on vacation? If it is enough,
then, it seems, we have an attractive externalist model of introspection. Just
as Clyde can know what it was Harold said without knowing, at least not
by hearing, that it was Harold who said it, why can’t a person know what it
is he thinks (by, say, introspection) without knowing, not by introspection,
that he thinks it?¹³

¹³ Strictly speaking, the analogy with knowing what you think vs. knowing that you
think it should contrast knowing what Harold said with knowing that he said it—not,
as I have done, with knowing that it was Harold who said it. With a few minor alterations
this could be done. All we need to imagine are things—programmed sound synthesizers,
for example, or (thanks to Doug MacLean for this suggestion) parrots—that can produce
the same sounds as Harold when he says that he is going on vacation without actually
saying or asserting anything. I assume here that parrots and machines who make the
sounds ‘I am going on vacation’ are not actually saying they are going on vacation. They
utter the words (and, therefore, perhaps, in direct discourse say) ‘I am going on vacation’,
but they do not, by producing these sounds, say (indirect discourse) that they are going
on vacation. When Clyde hears Harold on the phone saying that he is going on vacation,
therefore, he can know (by hearing) what Harold said without knowing (at least not by
hearing) that anything was said.
I have chosen to run the analogy as I have in the text because it is simpler and more
intuitive and it makes the point equally well. The important point, once again, is that
the way you know the x is y may be, and often is, quite different from the way you know
that it is x that is y.
80 Fred Dretske

I will be told (I have been told) by impatient skeptics that I am still

ignoring subtle, but nonetheless quite valid, distinctions of scope. Since
Clyde doesn’t know it was Harold who said he was going on vacation, it
would be wrong to put the phrase ‘what Harold said’ inside the that-clause
that expresses what Clyde knows. Clyde doesn’t know that what Harold
said is that he is going on vacation. All he really knows is something of or
about what Harold said—that it is that he is going on vacation. We can
say that Clyde knows what the caller said, and the caller was Harold, but
since Clyde doesn’t know the caller was Harold, he doesn’t know what Har-
old said. Clyde believes the caller was Harold, and he is (we may suppose)
fully justified in this belief, but (given the special circumstances) he doesn’t
know it. So it would be wrong to describe him as knowing that what Har-
old said is that he was going on vacation. Clyde doesn’t—not really, not
strictly—know what Harold said.
I don’t believe in this real, this strict, form of knowledge,¹⁴but I’m willing,
once again, to concede the point to those who believe that, strictly speak-
ing, nothing belongs inside the scope of a knowledge attribution that isn’t
known to belong there by the agent to whom the knowledge is ascribed. I
am, after all, interested in convincing even a stubborn internalist that there is
a perfectly workable externalist model of self-knowledge and that the teleo-
semanticist’s commitment to a historical account of thought and experience
is not to be rejected on epistemological grounds. So I’ll work with what I’m
given. My third (and final) example—a slight variation on the second—is
meant to comply with these more demanding strictures on scope.
We have to remember that we needn’t suppose that Clyde can know
what Harold said without knowing it was Harold who said it. All I need to
show is that the way Clyde knows it was Harold who said it can be different
from the way he knows what Harold said. If this is possible, then the way
is clear to concede (to the yet unconverted) that although you cannot know
what you think without knowing that you think it, the way you know what
you think may be entirely different from the way you know that you think
it. You can know what you think by introspection, but introspection may
not be the way you know you think it. To know that you are thinking about
pumpkins may require—or so externalists are free to maintain—a more
indirect method, a method (empirical investigation? historical research?)
compatible with a teleosemantic theory of what it takes to think about
pumpkins. All that introspection tells you is that it is pumpkins you are
thinking about.

¹⁴ Readers familiar with my views on contrastive statements (Dretske 1972), closure

(Dretske 1970, 1971, 2005) and the incremental character of perceptual claims (Dretske
1969, ch. ) will understand why I don’t accept it.
 Representation and Self-Knowledge 81

To illustrate this possibility we need imagine only a small modification

of our last example. Clyde (overly suspicious from too much philosophy)
finds out who called him by tracing the call. He discovers that the call ori-
ginated from a phone to which only Harold had access. So he knows it
was Harold who called him. He knows both what Harold said (namely,
that he was going on vacation) and that it was Harold who said it, but his
way of knowing the one is different from the way he knows the other. He
heard what Harold said, but he did not—indeed (given the impersonat-
ors) could not—have heard that it was Harold who was saying it. Given
the conditions, there is nothing distinctive about Harold’s voice to enable
one to know, by hearing him talk on the phone, that it is Harold. Clyde
knows it was Harold who said that he was going on vacation as a result of
an empirical research, but no investigation was required to find out what
Harold said. Clyde heard him say he was going on vacation.
Using this slightly modified example as a model for introspection, then,
the proposal is that the way we find out what it is we think (desire, wonder,
fear, expect, etc.) is different from the way we find out that we think
(desire, wonder, fear, and expect). The first method we call introspection.
Whatever, exactly, introspection comes down to, it does not involve
empirical investigation of external circumstances. That is what makes it
introspection. But this is quite consistent with an empirical investigation
being required to find out that the content revealed by introspection is the
content of a mental state, a state whose possession of content is constituted,
in part, by a network of external relations some of which are historical.
Given our model, this should be as sensible as saying that Clyde heard what
Harold said but needed an investigation (tracing the telephone call) to find
out it was Harold who said it. Or, to give an example that might appeal to
baseball fans, it is like needing a program to find out that it was Lou at bat,
but not needing a program to know that Lou hit a home run. You saw him
hit a home run. You know what he did by direct perception. But you know
who did it—that it was Lou—indirectly, by consulting your program.
We have, then, the following picture of self-knowledge: when think-
ing about or experiencing pumpkins, we can know, with a special kind of
first-person Cartesian authority, what it is we are thinking about and exper-
iencing—namely, orange pumpkins. Nothing illicit is smuggled into the
scope of the knowledge attribution since we know both that it is pump-
kins we are thinking about and that we are thinking about them. So we
can, in both word and thought, and with full knowledge, pick out and
refer to what we are thinking about—orange pumpkins—as something
we are thinking about. Nonetheless, our way of knowing that it is pump-
kins we are thinking about is, or may be, quite different from the way we
know that we are thinking about them. Although we enjoy first-person
82 Fred Dretske

authority about the first, we enjoy no privileged access to the second fact.
It may be, as teleosemanticists have it, that to know you are thinking about
or experiencing pumpkins requires information not obtained by looking
inward. Introspection doesn’t tell you that you think, only what you think.
If this is right, we have an answer to the epistemological objection to teleo-
semantics. A special authority about, and a privileged access to, one’s own
thoughts and experiences is compatible with a historical theory of thought
and experience. The only remaining question is whether this answer to the
objection gives a plausible account of self-knowledge. Is this really all that
introspection yields? Do we, in fact, use a different method to find out that
we think from the method (if it is a method) that tells us what we think?
My purpose here was only to argue that there was no valid epistemolo-
gical objection to teleosemantics. I’ve already done this. I should quit now.
But I can’t resist a few remarks about the plausibility of this picture of the
mind’s knowledge of itself.
As my examples show, we often know that x is y by some direct meth-
od (hearing, seeing, introspection) without knowing, without being able to
know, by that same direct method, that it is x that is y. If we know that it is x
that is y, our way of knowing this may be, and often is, quite different from
our way of knowing that x is y. I don’t have to see, even be able to see, that
it is water that is boiling to see that the water is boiling. There is, after all,
nothing about water to distinguish it from gin, vodka, and a variety of other
liquids. I needn’t be able to see that it is water in order to knowingly refer to
what I see to be boiling as water. So if it is water, and if I reasonably and truly
believe it is water, there is nothing to prevent me from saying I can see that
the water is boiling. That, I submit, is how I know that the water is boiling.
If I know it at all, though, that isn’t how I know it is water that is boiling. If
I actually know it is water, I probably know that in some way other than the
way I know it is boiling. The fact that I came to know (or believe) it is water
by chemical analysis doesn’t mean I can’t see that the water is boiling.
Why shouldn’t the same be true of introspection. The fact that I found
out I think by having someone (parents? teachers? friends? Descartes?) tell
me doesn’t mean I can’t now discover what I think by simple introspection.
The analogy with ordinary perception can be pushed a little further. Per-
ception of ordinary dry goods tells us what is in the physical world, not that
there is a physical world. I see that there are cookies in the jar, people in the
room, and (by the newspapers) continued violence in the Middle East. That
is how I know there are cookies, people, and violence in these places. Cook-
ies, people, and violence are physical things that exist independently of my
perception of them. Do I, therefore, know, by visual perception, by seeing,
that there are things that exist independently of my perception of them?
Can I see that there is a material world and that, therefore, solipsism is false?
 Representation and Self-Knowledge 83

I don’t think so. It seems more reasonable to say that assuming there is a
physical world, or assuming we know (in some other way) that there is a
physical world, perception tells us what sorts of things are in it—cookies,
people, and violence. Visual perception has the job of telling me what phys-
ical objects I see, not that I see physical objects. If my perceptual faculties
had the latter job, the job of telling me that I was (in effect) not hallucin-
ating, not aware of some figment of my own imagination, they would be
incapable of discharging their responsibilities. For, as we all know, hallucin-
atory cookie jars can, and sometimes do, look much the same as real cookie
jars. You can’t see the difference. If it’s a real object you see, perception
will tell you whether it’s an orange or a banana (a difference that is plainly
visible), but perception cannot tell you whether it’s a real orange or just a
figment of your imagination. That difference isn’t visible.
Memory has a similar structure. Memory tells us what happened in the
past—the specifics, as it were, of personal history. It does not tell us there
is a past. I can remember (hence, know) what I had for breakfast this morn-
ing. No trick at all. I distinctly remember that it was granola. Nonetheless,
despite this (what I remember) implying that the past is real (if it isn’t real,
I didn’t have breakfast this morning; hence, do not remember having gran-
ola for breakfast this morning) this doesn’t mean I can remember that the
past is real. If I know the past is real, I don’t know this by remembering
that it is real. That isn’t a way to answer Russell’s skeptical question about
the past.¹⁵ If I know the past is real, I know it in some way other than by
memory. Memory is a faculty that tells me what occurred in the past given
that there was a past just as perception tells me what is in the material world
given that there is a material world. Maybe I have to know the past is real in
order to remember what I had for breakfast this morning (I doubt it, but let
that pass), and maybe I have to know there is a physical world to see wheth-
er there are cookies in the jar (let that pass too), but the point is that I do
not have to know these things by memory and vision in order for memory
and vision to tell me (give me knowledge of ) what I had for breakfast this
morning and what is in the cookie jar.
Introspection is like that. Introspection tells me what is in my mind,
what it is I am thinking, wanting, hoping, expecting, and the kind of exper-
iences I am having. It doesn’t tell me I really have a mind, mental states
with content. If I know that at all, I know it in some way other than by
introspection, the faculty that, given that I have thoughts and feelings, tells
me what I’m thinking and feeling.

¹⁵ Russell’s question: How do you know the world and all its contents were not created
a few moments ago complete with memory traces, fossils, history books, etc.—complete,
that is, with all the indicators you rely on to tell you about the past?
84 Fred Dretske

REFERENCES

A, A., C, R., and P, M. (eds.) (2002), Functions: New Essays in
the Philosophy of Psychology and Biology (Oxford: Oxford University Press).
BË, S. E., and L, W. (1986), Knowing Who (Cambridge, Mass.: Bradford
Books, MIT Press).
B, C. (2002), ‘A Rebuttal on Functions’, in Ariew et al. (2002: 63–112).
C, R. (1975), ‘Functional Analysis’, Journal of Philosophy, 72/20: 741–65.
D, F. (1969), Seeing and Knowing (Chicago: University of Chicago Press).
(1970), ‘Epistemic Operators’, Journal of Philosophy, 68/24: 1007–23.
(1971), ‘Conclusive Reasons’, Australasian Journal of Philosophy, 49/1: 1–22.
(1972), ‘Contrastive Statements’, Philosophical Review, 81/4: 411–37.
(1981), Knowledge and the Flow of Information (Cambridge, Mass.; Bradford
Books, MIT Press).
(1988), Explaining Behavior (Cambridge, Mass.: Bradford Books, MIT Press).
(1995), Naturalizing the Mind (Cambridge, Mass.: Bradford Books, MIT
Press).
(2001), ‘Norms, History, and the Mental’, in Denis Walsh (ed.), Natural-
ism, Evolution and Mind (Cambridge: Cambridge University Press); previously
pub. in Dretske, Perception, Knowledge, and Belief: Selected Essays (Cambridge:
Cambridge University Press, 2000).
(2005), ‘The Case against Closure’, in Matthias Steup and Ernest Sosa (eds.),
Contemporary Debates in Epistemology (Malden, Mass.; Blackwell).
EÇ, B. (2002), ‘Indeterminacy of Function Attributions’, in Ariew et al. (2002:
291–313).
F, J. (1990), A Theory of Content and Other Essays (Cambridge, Mass.: Brad-
ford Books, MIT Press).
G, P. (1957), ‘Meaning’, Philosophical Review, 66: 377–88.
H, V. G. (2002), ‘On the Normativity of Functions’, in Ariew et al.
(2002: 144–56).
L, P., and M, N. (eds.) (1998), Externalism and Self-Knowledge (Stan-
ford, Calif.; CSLI Publications).
M, R. (1989), ‘In Defense of Proper Functions’, Philosophy of Science, 56/2:
288–302.
N, K. (1991), ‘The Teleological Notion of Function’, Australasian Journal
of Philosophy, 69: 454–68.
P, M. (2002), ‘Pagan Teleology: Adaptational Role and the Philosophy of
Mind’, in Ariew et al. (2002: 263–90).
W, D. M. (2002), ‘Brentano’s Chestnuts’, in Ariew et al. (2002: 314–37).
 4
The Epistemological Objection
to Opaque Teleological Theories
of Content
Frank Jackson

1. A COMMONPLACE ABOUT CONTENT

After you’ve seen someone walk through a minefield, you have a pretty
good idea where they think the mines are located. After you’ve dined reg-
ularly with someone, you have a pretty good idea of what they like to eat
and drink. And so on and so forth. It is a commonplace that what people
do and say tells the folk—that is, you and me, Shakespeare and Aristotle,
most of us when we are not drawing on specialist knowledge in cognit-
ive science or whatever—a good deal about the contents of the beliefs and
desires of our fellow human beings, and more generally about what they
think. The observation of behaviour in circumstances often grounds jus-
tified belief about the contents of intentional states, and, what is more,
we qua members of the folk know this. This commonplace does not pre-
sume behaviourism of course—we all know that the direction of the wind
is indicated by the behaviour of a windsock but facts about the wind cannot
be analysed in terms of facts about windsocks.
The fact that this is a commonplace means that any theory of content
should respect it. This chapter argues that certain versions of teleological
theories of content are inconsistent with it, and more generally with the fact
we folk qua folk have many justified beliefs about the contents of beliefs
and desires, and thought more generally. The versions I have in mind are
theories offering biconditionals like:
(A) x believes that P iff x is in a state that . . .
where the ellipsis is filled by a clause that relates to selectional matters
that are opaque to the folk, and where these biconditionals are understood
86 Frank Jackson

as part of a theory which identifies content properties with the relevant

selectional ones. The last proviso is important. Nothing I will say goes
against the views of those who are content to say that biconditionals like
(A) are interesting, or reflect something important about content, or are
helpful ways of understanding how it is we came to have contentful states
as we evolved, or . . . It is theories that identify content with opaque selec-
tional properties (or opaque teleological properties more generally, but we
will stick to selectional versions of teleology) that are my target. This is
not to make my target boringly small: surely the obvious way to under-
stand someone who offers a theory of content is as offering an identific-
ation of content properties, an account of what content is, and so the
obvious way to understand someone who offers a teleological theory of con-
tent is as offering an identification of content properties with selectional
properties.
An example of the kind of teleological theory that I am targeting is one
that says:¹
(B) x believes that P iff x is in a state that is selected (in the evolutionary
sense) to co-vary with P
and
(C) x desires that P iff x is in a state that is selected (in the evolutionary
sense) to bring about P
and sees these two clauses as making the case for identifying being a belief
that P with being selected to co-vary with P, for identifying believing that
P with being in a state that is selected to co-vary with P, for identify-
ing being a desire that P with being selected to bring about P, and for
identifying desiring that P with being in a state selected to co-vary with
P, or something along these broad lines. In sum, I mean a theory that
uses connections between selectional facts and intentional facts to deliver
identifications of being and having intentional states that P with being and
having certain selectional properties that have the appropriate relation to
that P.
Now the folk qua folk do not have opinions, let alone justified ones,
about the theory of evolution; and even if they did, they do not qua folk
have justified views about which structures are or are not selected for—it
took serious research to show that the chin is not selected for. Moreover,
even if the folk did have opinions including justified ones about which

¹ For a view of this kind, see Papineau (1993: 94). He is not offering it as a finished
theory but as a sketch to give the general idea. Our argumentation will be independent
of the various qualifications and refinements.
 Opaque Teleological Theories of Content 87

structures are selected for, they do not qua folk have opinions, let alone
justified ones, about what these structures are selected for —the selectional
histories of the ear and the larynx are major research topics. Moreover, no
one thinks that observations of subjects interacting with their environments
are enough to justify attributing properties like having a state selected to co-
vary with such and such, or having a state selected to bring about so and so,
to subjects. To suppose otherwise would make a nonsense of all the work
that went into establishing the theory of evolution.
At first glance, we seem to have a serious problem for teleological theories
of content of the kind in question, henceforth opaque selectional theor-
ies of content: they seem to imply that the folk do not have the justified
opinions about the contents of the intentional states of their fellow humans
that they clearly do have, and that, in particular, even extended observations
of interactions with environments are not enough to justify ascribing con-
tents. After all, the folk do not have justified opinions about magnolia metal
because they have never heard of it, and, even if they had heard of it, would
not normally have justified opinions about which metals are examples of
it.² In the same way, it seems that they would not, on opaque selectional
theories, have justified opinions about intentional contents because these
contents are properties they have never heard of, or need never have heard
of. And, even when they have heard of them, the absence or presence of
these properties is not something the folk need have a justified view about
in order to have justified opinions about what their fellows are thinking,
and is not something we get justified belief concerning out of observation of
interactions with the environment.
My contention in this chapter is that what seems right at first glance
seems right after a number of glances. I will argue that opaque selectional
theories of content cannot explain how it is that we folk have the justi-
fied opinions we do about intentional contents. They cannot explain, for
instance, why Shakespeare was so often justified in his views concerning
what those around him wanted and believed based on his observations of
their behaviour.
Our argument will be independent of the detail of the various teleo-
logical accounts of content provided only that they are opaque selectional
accounts in the sense already explained. Where necessary I will frame mat-
ters in terms of the very simple teleological account of the content of belief
and desire given above, but the point being made will not depend on the
simplification (which all parties can agree is substantial); it will depend on
the opacity.

² It is a lead-base alloy used in bearings.

88 Frank Jackson

2. OUR ARGUMENT LAID OUT

I know many teleological theorists will insist that their theory cannot be in
trouble from such a ‘quick’ objection and that the objection commits some
simple mistake or other. Most of the rest of this chapter is a series of replies
to the various objections to the folk epistemology objection, or FEO, as I’ll
call it, from people who insist that I have made some simple mistake or
other. But let me spell out the structure of the objection before we proceed
to look at some responses to it. The objection in schematic form runs thus:
Premise 1. Having an intentional state with such and such content is a
property justifiably ascribed in so and so circumstances. (Premise supported
by reflection on Shakespeare and the folk generally.)
Premise 2. Having a state, or being in a state, with so and so a selec-
tional history is not justifiably ascribed in so and so circumstances. (Premise
supported by reflection on what is needed to justify belief in selectional
matters.)
Conclusion. Having an intentional state with such and such content
is not identical with being in a state with so and so a selectional history.
(Leibniz’s Law)
This argument is valid and there is no problem about applying Leibniz’s
Law to properties of properties. The same style of argument can, of course,
be run for the intentional states themselves, for being the belief that P as
opposed to having the belief that P, for instance.
It is time to look at the possible objections. They are all constructed from
objections that I have come across in one form or another but I have not
sourced them for fear of misrepresenting.

3. THE OBJECTION THAT THE SELECTIONAL

WILL CORRELATE WITH THE FOLK

Suppose we have a theory of the form

(D) x believes that P iff x is in a state that is K
where K is an opaque selectional matter. We might have very good reason
to hold that
(E) x is in a state that is K iff x is in a state which is J
where J is a matter transparent to the folk. It will then be the case, according
to the theory, that
 Opaque Teleological Theories of Content 89

(F ) x believes that P iff x is in a state that is J.

For example, suppose that belief that P is a matter of being in a state
selected to co-vary with P but considerations to do, say, with the difficulties
of surviving should there be too much dissonance between what co-varies
with how things are currently and how things were as we evolved tell us
that current functional roles of the kind widely agreed to be transparent
to the folk will match up with the relevant selectional roles in ways that
mean the two are reliable guides to each other. In that case, runs the objec-
tion, it might well be that something like (F ) was true consistently with
content being determined by something like (D). In that case, continues
the objection, selectional theorists can escape the folk epistemology objec-
tion by pointing out that, although content is determined by something
opaque to the folk, the something opaque is reliably correlated with some-
thing transparent to the folk.
Reply. My response to this objection is that it is one thing for there to
be a correlation, another for the folk to be aware of it or to have a justi-
fied opinion that there is such a correlation. The history of medicine is full
of correlations we know about now and wish we had known about much
earlier—it would have saved many lives if we had known about them soon-
er. All the same, we did not know about them sooner and were not in any
position to have a justified opinion about them sooner. Perhaps, and any-
way I am granting the point here, the selectional roles are correlated with
matters transparent to the folk. This will not ground justified opinion on
the part of the folk about contents determined by opaque selectional facts if
the folk know nothing of the correlation and have no justified belief in its
existence. And isn’t this precisely the position of the folk? Shakespeare had
no idea, and could have had no idea, that there is a correlation between the
facts from which he justifiably inferred contents and the selectional facts of
teleological theory. But this did not debar him from having justified opin-
ions about what people were thinking.
The key point can be brought out by considering David Papineau’s dis-
cussion of an objection of Andrew Woodfield’s. Woodfield asks for a reason
to believe that ‘the teleological theory’s ascriptions of content coincide with
those made by everyday psychology’.³ Papineau’s answer ‘is simply that it
would be a mystery that the desire for some physical result r should do what
everyday psychology says it does [deliver what is desired for agents with true
beliefs] . . . unless it has been selected to produce r’.⁴ Be this right or not, it is
not something available to the folk. It would tell us why the folk get it right

³ Papineau (1993: 94).

⁴ Ibid. 97–8. He makes a similar argument for belief on p. 99.
90 Frank Jackson

given a selectional story but not why they are justified in holding that they
get it right.

4. THE OBJECTION FROM SCIENTIFIC IDENTITIES

Teleologists often present their view as a kind of scientific identification

of content properties and states with selectional properties and states play-
ing selectional roles.⁵ And it might well be thought that these identities
tell us that there must be something amiss with the FEO. The distribution
of water is not a separate matter from the distribution of H2 O, and yet
the folk can have justified opinions about where there is water despite not
having justified opinions about where there is H2 O, or even knowing that
there is such stuff as H2 O. Shakespeare had many justified opinions about
where there is water, despite not knowing any modern chemistry. Teleolo-
gists should, runs the objection, hold that as water is to H2 O, so content is
to selectional role.
Reply. It is important to appreciate what the water–H2 O example tells
us, and what it does not tell us.
It tells us that someone can consistently offer a claim of the form
(D) x believes that P iff x is in a state that is K

where K is an opaque selectional matter, and where (D) is advanced as a

necessary a posteriori truth and not as an a priori truth or as a conceptual
analysis in some traditional sense, without thereby committing themselves
to an epistemological claim about justified belief about content requiring
justified belief about opaque selectional matters.
What it does not tell us is the sense in which offering (D) is offering
what might properly be thought of as a teleological theory of content. As
it stands, (D) is compatible with theories of content that are antipathet-
ic to teleology. Consider the theory that holds that having a belief with
content that P, and being in a state that is K, are distinct properties, and
more generally that content is quite distinct from anything teleological: for
example, the theory might hold that being a belief that P and being selec-
ted to co-vary with P are quite distinct properties, and that believing that P
and being in a state selected to co-vary with P are quite distinct properties.

⁵ Papineau is explicit that this is how he understands teleological theories; see his
comments on theoretical reductions on (1993: 93), and his comments at the beginning
of (2001). Braddon-Mitchell and Jackson (1997) argues from the perspective of the critic
rather than the supporter, that this is the best way to read the view.
 Opaque Teleological Theories of Content 91

However, the theory also holds that although the relevant content prop-
erty and the relevant selectional property are distinct, they metaphysically
necessitate each other. This is not a teleological theory of content. It is no
more a teleological theory of content than are necessitarian dual attribute
theories of mind versions of physicalism. Some dual attribute theorists hold
that the phenomenal feels of sensory states are properties distinct from any
that appear in physicalist theories of mind, but that they are necessarily
connected to properties that appear in physicalist theories. The necessary
connection does not turn their view into a version of physicalism. Mutatis
mutandis for teleology.
It follows that we need to add something to (D) in order to get a view
that can properly be regarded as a teleological theory of content. The obvi-
ous addition, as we in effect noted near the beginning, is an identity claim.
In addition to holding that a suitable instance of (D) is a necessary a posteri-
ori biconditional, teleologists should hold that suitable instances of
(E ) having such and such content = being a state that plays so and so a
selectional role

and
(E ) being in a state with such and such content = being in a state that
plays so and so a selectional role

are necessary a posteriori truths. Again, the water–H2 O case might be held
to be suggestive. Not only is it a necessary a posteriori truth that x is water
iff x is H2 O, it is a necessary a posteriori truth that water is identical with
H2 O (modulo worlds where there is no water).
But now we have the trouble we noted at the outset. (E ) and (E ) are
false: for each it is the case that the property on the left-hand side differs in
its epistemic properties from the property on the right-hand side; that’s the
nub of the FEO. And a similar consideration applies in the case of scientific
identities. When Shakespeare believed that there is water in a glass in front
of him, he believed that things are a certain way in the glass, but what he
believed about how things are in the glass is not what we believe when we
believe that a glass contains H2 O. It follows that how things are believed to
be when water is believed to be somewhere is not how things are believed
to be when H2 O is believed to be somewhere. And this is what one would
expect given the a posteriori nature of the identity between water and H2 O.
If what one believes about how things are when one believes that there is
water is the same as what one believes about how things are when one
believes that there is H2 O, then from the very moment humans believed
that there is water, they believed that there is H2 O. But the a posteriori
92 Frank Jackson

nature of the identity between water and H2 O confirms the common-sense

view that believing that there is H2 O came later in time. The property
we ascribe when we believe that there is water differs from the property
we ascribe when we believe that there is H2 O. We believed the first to be
instantiated long before we believed the second to be instantiated. In sum,
if we call the first property ‘being water’ and the second ‘being H2 O’, there
is an epistemic argument that shows that being water  = being H2 O, just as
there is for having such and such content  = being a state that plays so and
so a selectional role, and for being in a state with such and such content  =
being in a state that plays so and so a selectional role.
Some will bite the bullet and say that we should think of the water–H2 O
case in the way many view the Hesperus–Phosphorus case. Many hold
that the proposition that Hesperus = Phosphorus is the same proposition
as that Hesperus = Hesperus, and, in consequence, that one believes the
first iff one believes the second, and that if one believes that some planet =
Hesperus, one believes that that planet = Phosphorus. Surprising but true.
In the same way, they might say, what we believe about how things are
when we believe that there is water is the same as what we believe about
how things are when we believe that there is H2 O, and, in particular, we
should say that Shakespeare believed that various glasses contain H2 O pre-
cisely because he believed that they contain water. Surprising but true. My
claim, therefore, that there is a difference in the epistemic properties of
being water and of being H2 O is a mistake, and ditto for selectional and
content properties.
However, the contention that the proposition that Hesperus = Phos-
phorus is the same proposition as that Hesperus = Hesperus is part of a
package deal that affirms that it is a priori that Hesperus = Phosphorus—
surprising but true—because we know that the one thing that those who so
contend should not say is that the proposition that Hesperus = Phosphorus
is not knowable a priori, whereas the proposition that Hesperus = Hes-
perus is knowable a priori. For ‘they’ are one and the same proposition on
the contention under discussion.⁶ But whatever should be said about the
thorny issues surrounding identities expressed using proper names, surely
scientific identities like ‘water = H2 O’ and ‘acids = proton donors’ are a
posteriori. (And teleologists who draw on scientific identities in explaining
their view insist on the point, see Papineau 1993).
I am not, of course, saying that water  = H2 O. The relation between
water and H2 O is like that between the colour of the sky and being blue.
The colour of the sky = being blue but the property we ascribe when we

⁶ For some of the issues here, see e.g. Soames 2002: 4 and the discussion that follows.
 Opaque Teleological Theories of Content 93

say that something has the colour of the sky is not being blue; it is being
same-coloured with the sky. Likewise, the property we believe something to
have when we believe it to be the colour of the sky is not being blue (on the
obvious reading); it is being same-coloured with the sky. Being water and
being H2 O differ just as being blue and being the colour of the sky differ,
but water and H2 O are one and the same just as (being) blue and the colour
of the sky are one and the same property.

5. THE OBJECTION FROM SECOND-ORDER PROPERTIES

Many of the things we say about how things are are to the effect that some-
thing has a property that itself has a property. To be fragile is to have
some internal nature that causes breaking on dropping: there is the inter-
nal nature, and there is what it does or would do. Or take the colour of the
sky example just mentioned: to say that something has the colour of the sky
is to say that it has a colour that has the property of including the sky in
its extension. We might call the properties ascribed in such cases ‘second-
order’, meaning not that they are properties of properties but that they are
properties possessed by x in virtue of x’s possession of a property that itself
has a property: the property of having whatever property is so and so. An
example much discussed in the philosophy of mind is the version of func-
tionalism tailored to be compatible with a type–type mind–brain identity
theory.⁷ It draws a sharp distinction between being in pain and pain. To
be in pain is to instantiate a kind that fills so and so functional role; pain is
the kind (one that may or may not vary from creature to creature, etc.) that
does fill the role. This gives two property identities:
being in pain = instantiating a property that fills so and so functional
role
pain = the property that fills so and so functional role (neural state N
in such and such creatures, as it might be).
In the same way, runs the objection to FEO, selectional theories must dis-
tinguish two properties. One is the property we ascribe to someone when
we say that they believe that P or (same thing) the property we believe them
to have when we believe that they believe that P. This property is not a
selectional property—it is, for instance, a property we folk are justified
in believing someone to have in circumstances where we are not justified
in believing them to have any relevant selectional property. However, this
property is the property of having a property which is thus and so, and this

⁷ See e.g. Lewis (1983) and Prior et al. (1982).

94 Frank Jackson

latter property is, as an a posteriori matter, a selectional property. We have,

that is to say, two property identities:
(F ) believing that P = the property of having a property that is thus
and so
(G) belief that P = the property that is thus and so.
The property referred to in the right-hand side of (F ) is the property we
believe that x has when we believe that x believes that P. It is the property
that we are entitled to ascribe on the basis of folk-available data including
especially interactions with the environment. Therefore, it is not a selec-
tional property. However, this does not preclude the property referred to
in the right-hand side of (G) being, as an a posteriori matter, a selectional
property. The objection from second-order properties to FEO holds that
the sense in which a selectional style of teleological theory of content is cor-
rect lies in the truth of (G), where the property referred to in the right-hand
side of (G) is indeed a selectional property.
Reply. There is a sense in which this is not a selectional theory of con-
tent. Content properties are the relevant commonalities among entities with
contentful states. The property of having a state with the content that P
is the relevant commonality among the creatures that have an intentional
state with content P. And the suggestion maintains that the relevant com-
monality for creatures that believe that P is having a property which is thus
and so, where this is not a selectional property. If it were, it would not
be a folk-available property. Similar remarks apply to the properties of the
intentional states per se. The point here is similar to the one often made
against the type–type mind–brain identity theory of mind. Although the
theory holds that pain is neural state kind N, the properly psychological
property is the relevant commonality among those in, say, pain, and that is
a functional property and not a neurological one on the theory. But let’s set
the branding issue aside. The substantive problem with the suggestion on
behalf of selectional theories that draws on second-order properties arises
from the need for the second-order property that figures in the right-hand
side of (F ) to be one we may reasonably ascribe on the basis of folk-available
data, to be one ‘we are entitled to ascribe on the basis of folk-available data
including especially interactions with the environment’.
The folk are not cognitive scientists. All the same, they know that when a
parcel of matter with clear boundaries displays distinctive, patterned inter-
actions with its environment, there are internal properties of that parcel that
play a big role in underpinning the interactions. It is very plausible that
when we believe that x believes that P, what we believe about x involves
some implicit assumptions about how the interactions are generated. This
 Opaque Teleological Theories of Content 95

is confirmed by the fact that when ‘blockhead’ cases and Martian marion-
ette cases are described to us, we immediately designate them as cases where
there is no contentful thought.⁸ That which generates ‘their’ interactions
is not of the right kind, and we know this qua member of the folk. It is,
therefore, plausible that the property we ascribe to x when we believe that x
believes that P, or when we use ‘believes that P’ of x, is a second-order prop-
erty in the requisite sense. But it is important that the ‘thus and so’ in (F ) be
given an undemanding reading in the sense of a reading that preserves the
point we have been focusing on: the folk qua folk have plenty of justified
beliefs about content. For example, if we spelt out ‘thus and so’ in a way
that stated that the internal workings be carbon-based, as in
(F carbon) believing that P = the property of having a carbon-based
property that is thus and so,

we would wrongly make it the case that the folk who do not know that
we are carbon-based—Shakespeare and Aristotle would be examples—lack
justified beliefs about content.⁹
All the same, this leaves quite a bit of room to manoeuvre. There is a lot
about internal goings on that is available to the folk from interaction pat-
terns. Consider the phenomenon of imprinting. Baby ducks are disposed
to keep company with the first thing of a suitable sort that they see after
hatching. Usually it is the mother duck obviously, but sometimes it is a
dog, the experimenter, or whatever. They imprint on the first thing that
they see, and their behaviour is explained in terms of their having imprin-
ted on the mother, the dog, or . . . Although we ascribe imprinting on the
basis of observation of behaviour, what we ascribe, or are in a position to
ascribe, goes well beyond behavioural patterns. We know, for instance, that:
the duck’s initial sighting lays down a persisting trace inside the chicken,
otherwise its following behaviour would fade away quickly; that the nature
of the internal trace that is laid down is a function of the nature of the
thing first seen, otherwise it would not be able to discriminate between the
thing first seen and things seen subsequently; and that the trace laid down
is causally connected to the ways the legs and the head operate, otherwise
the information being carried by the trace inside the baby duck would be

⁸ ‘Blockhead’ is Block’s (1981) example of a ‘person’ who works by look-up tree.

The Martian marionette is Peacocke’s (1983, final chapter) example of a ‘person’ who is
effectively a puppet controlled by radio signals from Mars.
⁹ This is not a quick argument for multiple realisability. Rigidifying is epistemically
undemanding—knowing that something is F is enough for knowing that it is actually
F. A suitable rigid reading of the ‘thus and so’ might, for example, reference-fix on
something which turns out to be carbon-based.
96 Frank Jackson

irrelevant to the movements of its head and legs in sustaining the accom-
panying behaviour. All of this is something available and implicitly known
to the folk. The little reasoning sketches given above do not call on spe-
cialist knowledge in cognitive science. However, it is also true that what is
so available is restricted to the causal, functional, and informational under-
pinnings of the behavioural interactions. It concerns leaving traces, causal
transactions between traces, casual links to legs, the causal role of the eyes,
and so on. In order to preserve the folk-availability of content, the ‘thus and
so’ in (F ) must pertain to matters of this kind. What is obvious to the folk
and something they may justifiably believe and ascribe on the basis inter alia
of observations of behaviour can be complex, but the complexity is restric-
ted to complex causal and functional roles, and the like. It does not include
selectional roles. If it did Wallace and Darwin’s insights would have been
available to the folk qua folk.
This blocks the possibility of the identification in (G) being with selec-
tional properties. The situation can be put as follows. We need an un-
demanding reading of the ‘thus and so’—one that makes the existence of
a property that is thus and so folk-available—in order for (F ) to meet the
folk-availability constraint. Mark any such reading ‘thus and so*’. But then
what we get from (G) is that belief that P = the property that is thus and
so* and the property that is thus and so* is not a selectional property. It
might be a functional property of the kind we can justifiably believe in
given the information available to the folk, or it might be whatever prop-
erty plays the relevant functional role, because if we can justifiably believe
that some functional property is instantiated we can justifiably believe that
there is a property playing the functional role. But in neither case is it a
selectional property. Selectional properties are not folk-available, and they
do not play the relevant functional roles—that is done by neural properties,
or maybe internal functional architecture.
In sum, the second-order property way of reading selectional cum tele-
ological theories of content faces a dilemma. Any reading of (F ) that allows
some instance of (G) to be an identification of a content property with a
selectional property is a reading that means that the folk-availability con-
straint is violated. Any reading of (F ) that meets the folk-availability clause
blocks any instance of (G) being an identification of content with a selec-
tional property.
 Opaque Teleological Theories of Content 97

6. THE OBJECTION THAT LEIBNIZ’S LAW DOES NOT

APPLY TO EPISTEMIC PROPERTIES

‘‘You went wrong near the beginning when you said ‘there is no problem
about applying Leibniz’s Law to properties of properties’. There is a prob-
lem if the properties of properties are epistemic ones like being justifiably
supposed to obtain in so and so circumstances. The opacity of belief con-
texts tells us that.’’
Reply. This is tantamount to denying that we have beliefs about prop-
erties. For consider the right response for those who hold that ‘a = b’, ‘S
believes that a is F ’, and ‘It is false that S believes that b is F ’ can be true
together. The right response is that when S believes that a is F, S does not
have a belief about a, the thing. S has, rather, a belief about the proposition
that a is F, and the explanation of how it can be false that S believes that
b is F and true that S believes that a is F, when a = b, is that the proposi-
tion that b is F is a different proposition from that a is F, and the beliefs in
question are about propositions and not objects. Thus there is no violation
of Leibniz’s Law. (Of course, some insist that S’s belief that a is F is about a,
but they are the same people who insist that if S believes that a is F, then S
ipso facto believes that b is F in the case where a = b.)
I think we should resist any suggestion that we do not have beliefs about
properties. We really do have beliefs about how things are in certain parts
of the world and how things are with certain things. Our beliefs really do
place things in categories, and that’s to assign them properties. Our assign-
ments may be correct or incorrect but it is properties, not something else or
nothing, that gets assigned.
It might be objected that an argument like the one just rehearsed for
objects can be developed for properties in order to show that we do not
strictly have beliefs about properties. Surely the following can be true
together
(H) S believes that a is blue,
(I) It is false that S believes that a has the colour of the sky,
( J) Blue = the colour of the sky.
How so if S’s belief is about the single property—blue (= the colour of
the sky)? The answer is that we can read (I) in two different ways. We can
read the claim that S believes that a has the colour of the sky as being true
when S believes that a is blue. Is there some other colour that S believes a
to have? On this reading it is not possible for (H), (I), and (J) to be true
together. The more obvious way to read (I) is as saying that it is false that
98 Frank Jackson

S believes that a has the property of being same-coloured with the sky. But
blue  = being same-coloured with the sky. There is no way to read (I) that
makes trouble for the intuitive view that when S believes that a is F, S has a
belief about a property, namely, the property that S believes a to have.

7. THE OBJECTION FROM THE REVISIONARY NATURE

OF OPAQUE SELECTIONAL THEORIES

‘‘To require selectional theories of content to meet the folk-availability

constraint is to misunderstand their revisionary nature. Yes, Aristotle
and Shakespeare lacked justified belief about what their fellow human
beings were thinking, but that is no objection to teleology of the opaque
selectional kind once its revisionary nature is taken into account.’’
Reply. There are revisionary theories and there are revisionary theories.
In the company of many, I think that there is a revisionary element in
many of the most widely discussed analyses in philosophy.¹⁰ Compatibil-
ist analyses of free will and continuity accounts of personal identity often
have a normative element; they are not purely descriptive. ‘‘This is what
our concept should be when we take into account all we know, identify
potential confusions in the folk concept, and the role we give the concept’’
is the tenor of the discussion. However, these analyses preserve our epi-
stemic entitlements. They do not imply across the board that our beliefs
about which actions are free and when the person I am talking to today is
the person I was talking to yesterday are ones we are not entitled to. For
example, continuity theorists who allow that there is a Cartesian element
in the folk concept of personal identity allow that Shakespeare was by and
large entitled to his opinions about when he was or was not using the same
actor to play different parts. Indeed, often the reason for favouring some
analysis or other of personal identity or free action is in part that it makes
sense of how it is that our judgements about identity and free action are
reasonable ones.

REFERENCES

B, N (1981), ‘Psychologism and Behaviorism’, Philosophical Review, 90:

5–43.
B-M, D, and J, F (1997), ‘The Teleological The-
ory of Content’, Australasian Journal of Philosophy, 75/4: 474–89.

¹⁰ See Jackson (1998: 42–3).

 Opaque Teleological Theories of Content 99

J, F (1998), From Metaphysics to Ethics (Oxford: Clarendon Press).

L, D (1983), ‘An Argument for the Identity Theory’, repr. in Lewis,
Philosophical Papers, i (New York: Oxford University Press), 99–107.
P, D (1993), Philosophical Naturalism (Oxford: Blackwell).
(2001), ‘The Status of Teleosemantics, or How to Stop Worrying about
Swampman’, Australasian Journal of Philosophy, 79/2: 279–89.
P, C (1983), Sense and Content (Oxford: Oxford University
Press).
P, E. W., P, R, and J, F (1982), ‘Functionalism
and Type–Type Identity Theories’, Philosophical Studies, 42: 209–25.
S, S (2002), Beyond Rigidity (New York: Oxford University Press).
 5
Useless Content
Ruth Millikan

A central claim of teleological theories of content, certainly of my own

teleological theory of content, has been that the content of a representa-
tion is determined, in a very important part, by the systems that interpret
it. I should be more careful: . . . by the interpreting systems with which the
representation-producing systems have been designed to cooperate. Con-
tent is determined, in large part, by the ways these interpreting systems are
designed to react to the representations presented to them. More exactly,
content is determined by the rules or functions in accordance with which
representations need to correspond to world affairs (to what they represent)
if the responses of these interpreters are going to produce the effects they are
designed to produce by way of normal mechanisms (LTOBC, VM ).¹
Teleosemantics is in large part ‘consumer semantics’. It cannot be the
function of a system to produce representations unless the representations
that are produced have some use. But they will not have a use unless they
have consumers that know how to use them, consumers such as, perhaps,
some further systems in the brain or in the brains of conspecifics. Further,
if design is interpreted to mean biological design and use is interpreted to
mean biological function, then it looks as though the content of a repres-
entation must depend, in the end, on its having biological utility.
Various untoward consequences have been claimed to follow from this
view. I will discuss some of these alleged consequences. Most of what I
will say I have said before one place or another, but in diverse contexts
in a variety of different essays. I will try to draw these various thoughts
together.

The first worry is sometimes raised in a broader context. Nonhuman

animals, it is said, mentally represent only those aspects of the world that

¹ I use the following abbreviations: LTOBC for Language, Thought and Other Biological
Categories; OCCI for On Clear and Confused Ideas; VM for Varieties of Meaning.
 Useless Content 101

they need to represent in order to guide their limited repertoires of beha-

vior. The worlds that they live in are severely restricted; they register only
those aspects that are directly relevant to fulfilling their practical needs.
Add to this the increasingly popular Gibsonian position that perception
is always, at root, perception of affordances. The animal perceives its envir-
onment only as a series of possibilities for action, as containing things to
approach or to run from, things to climb up on, places to go through,
things to eat, things to hide under, things to throw, and so forth. The anim-
al’s world is perceived only as it relates directly to its own interests, needs,
and abilities. What is afforded to an animal is necessarily a function of its
own peculiar capacities for action. Where animals have different kinds of
abilities, then, they must live in different perceived worlds. Nor did Gib-
son restrict this description of perception to nonhuman animals. Human
perception is included. Apples are perceived as for eating, mailboxes as for
posting letters in, and so forth. But if this is the character of our basic per-
ception, and if human cognition is built on top of equipment originally
designed for this kind of perception, it seems to follow that we too live in
a world with subjectively imposed limits. Our world cannot be an objective
world, but is a world created in large part by our peculiar human con-
stitutions and abilities. The view that the content of every representation
depends in an essential way on its use seems to play directly into the hands
of this kind of argument. If teleological theories of content are right, we
humans should be intrinsically unable to represent a truly objective reality.
The enterprise of empirical science is doomed to presenting only a warped
and truncated vision of the world.
I agree with Gibson that basic perception is perception for action, indeed,
that basic perception is perception of affordances (VM, pt. ). Gibson’s
position well deserves to be gaining in popularity. But for reasons I will
review below, not all human perception is perception directly for action,
nor is all cognition based on perception designed directly for use in action. I
will return to this theme a bit later. At the moment there are more element-
ary points to be made.
First, a perception or other representation of the relation of something to
oneself or of the utility of something for oneself is not the same as a subjective
representation or a representation of something merely subjective. It may be,
of course, that when the objects one observes have certain kinds of relations to
oneself, utilities or disutilities for oneself, then biases in observation are more
likely to occur. But affairs involving relations to oneself or utilities for oneself
are, in themselves, perfectly objective affairs. There is no intrinsic reason why
an animal that perceives only affairs of practical significance to itself should
not perceive these affairs perfectly objectively and perfectly accurately. For
example, in order to manipulate any object in my world I will need to perceive
102 Ruth Millikan

its spatial relation to me. But its spatial relation to me is a perfectly objective
relation, and I am best off if I perceive it accurately.
Second, that a creature represents only a few features of its world does
not mean that it represents its world as having only a few features. The limit
of one’s representations of the world is not a representation of the limits
of one’s world. That two creatures represent different aspects of the world
they live in does not imply that they represent different worlds or that they
represent the world as being different ways. What I know about the world is
very strictly limited. It does not follow that I represent the world as having
strict limits. In the vocabulary of OCCI, to suppose that this follows would
be to ‘import completeness’ (OCCI, ch. 8).

A second worry about how the content of a representation could rest on

its biological utility may be eased by understanding clearly the breadth of
the notion of biological utility that is intended. The claim is that all human
purposes may be seen to be biological purposes when examined in the right
light, hence that all human utility is biological utility (VM, ch. 1). Hav-
ing biological utility is not at all the same as merely serving a purpose for
which the organism’s genes were selected. Besides biological purposes that
rest on genic selection, such as the purposes of the stomach, heart, liver,
kidneys, the purposes of their various activities, and so forth, instrumental
conditioning generates a separate level of purposive activity, and explicitly
represented goals implemented through practical reasoning yet another.
The relations among these various levels of purpose is complicated. I will
say something about this.
One of the things that has been evolved by natural selection is evolv-
ability itself, for example, in the evolution of homeo box genes and of sexual
reproduction. Relevant to our current concerns is the evolution of behavi-
oral systems that learn by trial and error or, as Dennett says, by generate
and test. The analogy between genic selection and operant conditioning
or instrumental learning is remarkably close (Hull et al. 2001). Processes
of practical reasoning also proceed by trial and error, as the thinker men-
tally experiments, attempting different routes from his current situation to
another envisioned situation until a path is found that will connect these
two (VM, ch. 17). Different kinds of purposes emerge on each of these
levels. These purposes sometimes conflict with one another and also with
the original purposes for which the genes were selected, yet all are biolo-
gical purposes, originally derived from natural selection of genetic materials
(VM, chs. 1 and 2).

Consider first instrumental learning. Appropriate genes are responsible

for the development of the systems that learn by operant conditioning. And
 Useless Content 103

appropriate genes are responsible for determining what the original forms of
reinforcement will be for a particular animal species. But forms of reinforce-
ment, such as alleviation of hungry or of thirsty feelings, or the presence of a
sweet taste in the mouth, are not themselves means toward fulfillment of any
purposes of the genes. Sweet tastes, for example, although correlated during
the history of the species with the presence of needed nutrition, are not, in
themselves, involved in any direct causal process resulting in increased nutri-
tion. Behaviors reinforced by sweet tastes (M&Ms) are selected for
producing sweet tastes, so they have as a natural purpose to bring in sweet
tastes. Also, true, the disposition to be reinforced by sweet tastes was selec-
ted because sweet tastes typically indicated nutritive value. But behaviors can
succeed in bringing in sweet tastes without succeeding in increasing nutri-
tion. That is what saccharine is sold for. Although the purpose of procuring
sweet tastes on the genetic level is gaining calories, the psychological purpose
of behaviors conditioned by sweets is merely to procure sweet tastes. So beha-
viors can succeed in fulfilling their natural biological purposes of bringing
in sweet tastes without fulfilling any more basic functions for which genes
were selected. In the broad sense meant here these behaviors have ‘biological
utility’, although in a narrower sense—on a shallower level—of course they
do not. We generally distinguish these levels of purpose by calling one ‘psy-
chological’ and the other ‘biological’, but looked at carefully, the level of
psychological purposes produced by reinforcement is just another layer of
biological purpose, though these are not so direct as the purposes, say, of
the heart’s beating or the stomach’s juices flowing. It is in the broad sense in
which psychological purposes are included among biological purposes that
the content of a representation is said to rest on its biological utility.
Processes of practical reasoning also proceed by trial and error, generat-
ing a third level of natural selection, hence of natural purposes. The nat-
ural purpose of behaviors selected through practical reasoning is to reach
whatever goals are represented at the start of the reasoning as the ends for
which means are to be selected. The final implementing intentions that
immediately generate behaviors have been selected because these repres-
entations have led, in the inner world of the thinker, to representations of
the original goal as fulfilled. If the practical reasoning operates properly,
implementing these selected intentions will produce implementation of the
original goal. The interesting and challenging question concerns the origin
of these represented goals, in particular, their relation to more basic biolo-
gical purposes resting more directly on genetic selection. What determines
which conscious goals we pursue? Perhaps most are derivative from prior
goals that we have. We aim for them because we believe they will lead to
ends already established as goals. But what mechanism selected the original
goals from which these goals were derived?
104 Ruth Millikan

Clearly our original conscious goals are not the same as whatever our
genes aim for. Babies do not come into the world fervently desiring to
live a long life and produce lots more babies. Nor do they come into the
world desiring to obtain just those goodies that will or would serve to
reinforce their behaviors through conditioning. This is true in the first
instance simply because babies don’t come into the world thinking about
anything, hence explicitly desiring anything. They have to develop concepts
before they can explicitly desire things. But let us suppose that a child has
developed adequate concepts of all the things that do or will happen to
reinforce her behaviors, from sweet tastes to sexual pleasure to smiles. She
is able to think about each of these things. Merely developing concepts
of these things is not the same as conceiving these things AS things that
reinforce or AS things wanted, however. Just as there may be a large gap
between what reinforces an animal’s behavior and what has deeper biolo-
gical utility for the animal—between sweet tastes and nutrition or between
sexual pleasure and having babies—there may be a large gap between what
actually reinforces behavior and what one wants or what one understands
to be a cause of reinforcement. Often we may know when we are attracted
or repulsed without knowing exactly why we are attracted or repulsed, or
we may know that we are happy or sad without knowing why. Knowing
exactly what it is that we want is by no means automatic. Likely we only
find this out by experience, indeed, by something analogous to hypothesis
formation and testing. ‘I don’t know why it is but shopping for Christmas
always makes me anxious’, we say, or ‘There is something about electric
trains that just fascinates me.’ We may wrongly suppose that having a lot
of money is what is required to make us happy, or being allowed to sleep
in every day without having morning commitments. Indeed, just what it is
that makes people happy has proved to be an extremely challenging ques-
tion for clinical and social psychology. Moreover, knowing what it is that
attracts us or repels us need not lead to a reasoned desire for or against
that thing. Often we have other conflicting interests to consider. Still, the
mechanisms through which goals are projected by conscious desire and
reason are undoubtedly mechanisms that our genes have been selected for
engendering.² These mechanisms are still with us because they have some-
times—often enough—produced behaviors that benefited human genes in
the past. Aims and goals that are products of these mechanisms are also bio-
logical purposes in the broad sense intended, even though it may happen
very often that they do not lead to fulfillment of any more direct purposes

² The claim, made by Fodor and others, that our human cognitive mechanisms may
actually have arrived of a piece without benefit of natural selection is discussed in VM,
ch. 1 n. 2 and ch. 2 n. 5.
 Useless Content 105

of the genes, perhaps not even producing behaviors that are psychologically
rewarding. Fulfillment of these aims and goals still has biological utility in
the broad sense that is meant.
Now one of the most important jobs that beliefs are designed to do,
surely, is to combine with other beliefs to form new true beliefs. This kind
of function contributes to biological utility whenever any of the beliefs
formed in this way turns out to have biological utility. Even the most highly
theoretical of beliefs are not excluded from having biological utility, then,
so long as they participate in chains of reasoning that eventually bear prac-
tical fruit. Still, surely many beliefs that we humans have do not ever help to
serve any of our goals, so the question does arise how these beliefs can have
content on the teleologist’s view.

First we should be clear that the teleologist’s position is not that each indi-
vidual belief must help to serve a biological function. Rather, beliefs must
fall within a general system of representation where the semantic rules for
the system are determined by reference to the way its consumers, its inter-
preters, are designed to use these representations. But in order to have been
designed to make and use representations in the system in a certain way,
the producers and consumers or their ancestors must have been using rep-
resentations from the same system productively in the past. Otherwise these
mechanisms would not be representation producers and consumers. They
would not have been selected, through evolutionary history or through
prior learning, for their capacity to coordinate through the use of repres-
entations. But there is another alternative. There may be ways that the
producers have been designed to learn or to be tuned to produce repres-
entations that are coordinated with ways the consumers are designed to
be tuned to use them, so that the producer’s rules and the consumer’s
use dispositions are somehow tailored in advance to match one another.
If the former—if the producer and consumer (or their ancestors) have a
history of being coordinated through the use of a particular semantic sys-
tem, we need to know how to generalize from past successful coordinations
to determine a unique general semantic rule that determines content in
those cases where no actual coordination occurs. If the latter—if produ-
cer and consumer have been designed to learn to respond in a coordinated
way without actually practicing together, we need to understand how this
marvel is accomplished. I have argued that the semantic rules governing
representations that humans use in perception and thought are sometimes
derived one of these ways and sometimes the other. Let me discuss these
alternatives in turn.
Suppose, first, that the rules are determined by a series of past successful
coordinations between the producers and consumers of the representations.
106 Ruth Millikan

These mechanisms were selected for as a result of these coordinations, selec-

ted for either by learning systems or through genetic evolution. The semant-
ic rules for these representations concern mapping relations between them
and world affairs they have corresponded to such that two conditions are
fulfilled. First, there must be a single kind of causal mechanism or process
that operated in these historical cases, reference to which explains how the
producers managed on each of these past occasions to produce representa-
tions that corresponded to world affairs by these mapping rules. It cannot
be the proper function of the producers to make representations that vary
in a definite way parallel to affairs obtaining in the world unless there is a
systematic way they have been doing this under certain conditions. They
must, for example, be sensitive to ‘local natural signs’ of the existence of
these affairs (VM, chs. 3–6). Second, there must be uniform explanatory
causal mechanisms or processes that have accounted for the fact that the
consumers’ responses to the representations have contributed in the past to
proper performance of their functions and these explanatory mechanisms
must depend on the fact that the representations have varied according to
world affairs by some particular semantic rule of correspondence. Repres-
entations that are not actually used have content because their semantics
is ‘compositional’ in a very broad sense (Millikan 1995, 2004). Complete
representations always fall within some system of representations such that
variations in the complete representations correspond systematically to vari-
ations in what is represented.³ The content of representations that are not
used or not used productively is determined by the way other representa-
tions in the same system have been successfully used in the past.

Still, we might question whether there is only one way to generalize from a
series of past successful uses so as to determine a general rule of correspond-
ence between representations and candidates represented. Recall Kripke’s
(1982) worries that no matter how many past examples of correct addi-
tions one begins with, these examples will never determine the correct way
to go on to new examples using the ‘plus rule’. But the two cases are very
different. The rule we are seeking is one that not only coincides with past
successful cases of coordination but conformity with which, in each case,
causally generated the coordination. In each case, had the representation
been different the consumer would have reacted differently, and this reac-
tion would not have served its purpose, or would not have served it owing
(in part) to the fact or condition that was represented. Kripke’s difficulty
was that he could not appeal to dispositions that he has to react to the plus

³ For a discussion of this sense of ‘compositional’ and its application both to

intentional representations and to natural signs, see VM, chs. 3–4.
 Useless Content 107

sign, because under certain conditions or at certain times his disposition

may be to make mistakes. But we can appeal to dispositions. We can appeal
to the dispositions that past representation consumers have had at those
particular times and in those particular circumstances in which their reac-
tions to representations served their psychological or biological functions.
The particular dispositions to which we appeal are relational dispositions.
How the properly functioning consumer reacts is determined as a func-
tion of its representational input. Its disposition is to react in a way that
bears a certain relation to this input. Different inputs would have produced
different outputs. Kripke’s deep difficulty was that he could find no way
of determining objectively when the attempt at addition gives a correct
answer. Indeed, although he doesn’t mention this, he has no way of determ-
ining which past additions were correct, hence were suitable examples from
which to generalize either. But we do have a way of determining object-
ively when representation production and consumption have achieved a
coordination in the past.

The skeptical question can be pressed further, however. Representations

that are not in fact used to serve biological utilities are one thing. But could
a representation have a semantic value that was so extreme that, supposing
it to be true, it would be physically impossible for it’s consumers to make
use of it. For example, Caroline Price (2000) asks about a bee dance that
indicates nectar in a position that is far too distant for any bee to fly to.
The bees do not, presumably, use the information gained from the dances
of fellow bees in theoretical inference. The semantics of bee dances derives
solely and directly from the use of these dances in directing bees to fly to
sources of nectar. Now in point of fact, bees indicate greater distance by
slowing down their tail waggles and lengthening the intervals between the
bursts of sound they make while dancing. Besides the fact that no bee could
return from a distance that was too far to fly to in order to perform a dance
indicating this distance, it is also unlikely that bees have the capacity either
to dance properly going so slowly or to perceive accurately the speed of wag-
gling and of sound bursts below a certain frequency. For what would they
use these capacities? It is unlikely that a dance that, by logical extension of
beemese rules, would tell of nectar much too far to fly to could be either
danced or, more central, recognized by fellow bees. No ancestor bees have
had dispositions to make use of such dances. Such bee dances, then, are
meaningless in beemese. After all, bee dances are not designed for human
interpreters who carry stopwatches to read them, but for bees. Turning to
humans, my claim is that the only human mental representations whose
semantic values are determined solely by generalization from past successful
uses are those perceptual representations that directly guide action, such as
108 Ruth Millikan

are produced, for example, through the dorsal visual channels (VM, ch. 14).
And there are surely limitations on the extremes that humans can perceptu-
ally represent or interpret directly for action, just as there are on the dances
that bees can dance and interpret.
Pressing questions about the representation of affairs with which humans
could not possibly interact arise only when we are pretty certain that we
do in fact represent these things. For example, Peacocke (1992) has asked
about representing propositions that concern things outside our light cone.
No interaction with such things is physically possible even in principle.
What makes this question legitimate is that we are pretty sure that we
can formulate questions about what is outside our light cone, and even
though we are not able to answer these questions, it is possible that someone
might have an unfounded belief about some such matter, and that the belief
would in fact have a truth value.
Again, my suggestion on this matter will involve an appeal to composi-
tionality.⁴ But first, I need to locate the problem within the framework of
the theory of empirical concepts with which I have been working (LTOBC,
OCCI ). Representations produced and consumed by systems that employ
empirical concepts are examples of a kind of representation that producers
and consumers can learn how or be tuned to use cooperatively without
actually practicing together. Their production and use dispositions can be
tailored in advance to fit one another. It will take me a few moments to
explain how this can be. I will not defend my position on this matter at any
length, however. Discussion and defense can be found in VM chapter 19,
LTOBC chapters 15–19, and OCCI chapter 7. The implied realist onto-
logy is explained and defended in OCCI chapter 2 and in LTOBC chapters
14–17.

Empirical concepts, in basic cases, consist in part of abilities to recognize

the objects, kinds, properties, and relations (I will just say entities) of which
they are the concepts when these entities are encountered in experience.⁵
Adequate abilities of this sort must be extremely sophisticated abilities. This
is because the same entity encountered under different conditions, at differ-
ent angles and distances from the perceiver or registered through different
media of information transmission, will impact the perceiver through a

⁴ So does Caroline Price’s solution (2000). I agree with Price’s analysis as far as it goes.
But the question how we acquire concepts of objective kinds, objects, and properties in
the first place so as to recombine them compositionally requires explanation.
⁵ In OCCI, ch. 6, and in VM, ch. 9, I argue that gathering information by believing
what another human says is in all relevant ways exactly like gathering information
through direct perception. The result is that people can have basic concepts of things
that they don’t yet know how to recognize ‘in the flesh’.
 Useless Content 109

wide variety of different proximal stimulations. The most formidable of all

the tasks that the perceptual–cognitive systems must perform is to learn to
recognize what is objectively the same as the same when it is encountered
again under any of numerous varying circumstances. Lacking this ability,
an inner representation-producing mechanism would either fail to recog-
nize the presence of its represented objects nearly all of the time, or be
ridiculously redundant, producing many separate and different represent-
ations of the same, which the consuming systems would not then recognize
as being representations of the same. This would result, first, in an inability
to accumulate information about an entity over time in a format allow-
ing recognition that the information all concerned the same entity. Second,
it would result in an inability to apply information previously gathered
about entities when interacting with them again, for these entities would
not be recognized as the same ones the information concerned. Having an
adequate ability to recognize when one is encountering information about
the same thing again is an essential part of what it is to have an adequate
empirical concept of that thing. Assuming that we are not born with con-
ceptual abilities of the required sort, how do our perceptual–cognitive sys-
tems learn to reidentify objects, kinds, and properties correctly? What cri-
teria do they use to distinguish successful from unsuccessful attempts at
reidentification so as to fashion a nonredundant (and nonequivocal) inner
representational system capable of representing these entities and their con-
figurations in a uniform way?

A simple reply is that correct reidentification of these entities might be

tested during practical activity. What counts as the same is what yields the
same results when reacted to or treated in the same way. It is the same, for
example, if you can chase and catch it using the same technique, and eat
it in the same way, and if it tastes the same way, and nourishes you in the
same way. It is the same if in responding to it the same way you get the same
proximal results. This answer assumes that you can recognize when your
responses or treatment are objectively the same, but although attempting to
stabilize two variables at once makes things more difficult, there is no reason
in principle why it can’t be done. No doubt this is the way in which natur-
al selection trains the innate perceptual–cognitive mechanisms of a species
and how perceptual–cognitive learning takes place both in lower species
and, at a certain level, in humans. Psychologists call this ‘generalization and
discrimination’, though it is not always recognized that learning what is
objectively the same response, with regard to the practical purposes at hand,
is part of what is being learned.
But this simple reply leads us directly back to the very first problem
raised in this chapter. If the only criterion for having successfully learned
110 Ruth Millikan

how to reidentify an object, kind, or property is successful use of that abil-

ity in practical activity, then one could only learn to represent those aspects
of the world one needed to represent in order to guide one’s behaviors.
Elsewhere I have argued that it is likely that the representational abilities
of nonhuman animals are limited in exactly that way (VM, chs. 18–19).
But there seems reason to believe that human cognitive capacities are not
so limited. Is there a way that we humans might learn to identify object-
ive entities with which we have had, as yet, no practical dealings, so as to
accumulate information involving them within a non-redundant represent-
ational system? If we could develop abilities accurately to reidentify various
entities in the world prior to practical uses, we could use these abilities later,
much more easily learning to use these entities in a variety of practical ways
and to apply prior knowledge of them during practical activities. A general
ability to discover where objective distal samenesses lie through the inter-
vening diversity of proximal stimulations would be selected for because it
served further biological functions. It would not then be necessary that
every conceptual ability developed in this way should actually find prac-
tical applications. The belief representations in which a concept that had
been developed in this way came to figure would have determinate truth
conditions without that.
The suggestion on conceptual development that I have made (LTO-
BC, chs. 18 and 19; OCCI, ch. 7; VM, ch. 19) might be summed up
in old-fashioned terminology by saying that coherence in human belief
serves as a test of correspondence. It serves as a test of the correspond-
ence, not primarily of individual beliefs to their truth makers but, more
fundamentally, as a test of consistent correspondence of the same represent-
ational elements to the same elements in objective world affairs. Coherence
serves as a test of our capacities correctly to reidentify distal entities through
diverse proximal manifestations. Consistent judgment is a natural sign of
the representation producer’s capacity to represent the same as the same.
Coherence is a psychological goal that reinforces ways that the cognitive
systems are producing representations. Like a sweet taste in the mouth,
coherence is a sign that has correlated with achievement of a deeper biolo-
gical purpose, the production of a consistent and accurate representational
system, itself only a means to further biological ends. Now I will unpack
this a bit.

Consistency within a representational system can be used as a test only

if inconsistency is possible within that system and if inconsistency shows
up on the surface of the representational system. Consider bee dances, for
example. Bee dances cannot conflict with one another. If two bees dance
different dances when returning to the hive, it is always possible that both
 Useless Content 111

dances are correct. There really is nectar both of those places. Nor do
the bees have any way of representing where there isn’t any nectar. Bee
dances are not sensitive to a negation transformation. A subject–predicate
sentence and its negation, on the other hand, are explicitly incompatible,
incompatible right on the surface. Similarly, humans can think negative
thoughts, and these thoughts contrast explicitly with possible positive
thoughts. Whether the way human thoughts are coded resembles the way
language is coded in any other way, certainly our thoughts are sensitive
to a negation transformation. Whenever we have opportunity to gather
the same information in two different ways, through two different natural
information channels yielding different proximal stimulations, we have a
chance to gain evidence about whether our various methods of attempting
to identify the entities represented in the subjects and the predicates of
these judgments are each converging to focus on some single objectively
same thing. Consistent agreement in judgments is evidence that these
various methods of making the same judgment are all converging on
the same distal affair, bouncing off the same target, as it were. If the
same belief is confirmed by sight, by touch, by hearing, by testimony,
by various inductions one has made, and is confirmed also by theoretical
considerations (inference is a method of identification too), this is sterling
evidence for the univocity of the various methods one has used to identify
each of the various facets of the world that the belief concerns.
Thus the same object that is square as perceived from here should be
square as perceived from there and square by feel and square by check-
ing with a carpenter’s square and square by measuring its diagonals and
square by hearing from another person that it is square.⁶ Similarly, if a per-
son is tall and good at mathematics as recognized today, that same person
should prove tall and good at mathematics when reidentified tomorrow.
Both one’s general methods of reidentifying individuals and one’s methods
of recognizing height and mathematical skill are corroborated in this way as
methods of reidentifying objective selfsames. That the same chemical sub-
stance is found to melt at the same temperature by checking with an alcohol
thermometer, a mercury thermometer, a resistance thermometer, a gas ther-
mometer, a bimetal expansion thermometer, and a thermal thermometer
is evidence both that one is able to recognize the same chemical substance
again and that there is indeed some real quantity (unlike caloric pressure)
that is being measured by all of these instruments. If a multitude of differ-
ent operational definitions are found to correlate with one another exactly,
then they can be assumed all to measure the same thing. Moreover, and

⁶ See n. 5 above.
112 Ruth Millikan

much more fundamentally, it can be assumed that there is something real

that each one measures.⁷

Notice, however, that agreement in judgments can be a test of correspond-

ence only in so far as it is possible for one’s judgments to disagree with
each other. And explicit disagreement in judgments is possible only in so
far as negation in judgments is possible.⁸ In its basic form, negation is a
semantic operation on the logical predicate of a sentence (Millikan 1984,
ch. 14; Horn 1989, ch. 6).⁹ Logicians call this ‘internal negation’. For
example, the normal reading, say, of the classic negative sentence ‘The
king of France is not bald’ makes it equivalent to ‘The king of France is
non-bald’, so that the negative as well as the affirmative presupposes the
existence of a king of France. More obviously, ‘John is not tall’ is equival-
ent to ‘John is non-tall’ and ‘John does not know French’ is equivalent to
‘John is ignorant of French’, and so forth. There are also secondary uses
of ‘not’ to reject a sentence on non-truth-conditional grounds, as in ‘The
slithy toves did not gyre and gimble in the wabe’ or ‘The square root of two
is not blue’ or ‘You didn’t see two mongeese, dear, you saw two mongooses’
or ‘The king of France is not bald, dear; France doesn’t have a king’. But
the fundamental use of the negative is not to prohibit assertion of a sen-
tence, but to make a positive, though indefinite, statement to the contrary.
The standard negative sentence actually says something positive about its
subject, namely, that it is characterized by some contrary or other of the
predicate of the sentence. If John is not tall it is because he is short or of
medium height. If John does not understand French it is because French
sentences either leave his mind blank or produce in it thoughts different
than for a Frenchman.

This point about negation assumes importance when we turn to epistemo-

logy and consider how evidence is gathered for a negative judgment. Begin
with the obvious: The absence of a representation of a certain fact is not

⁷ ‘If we see on a road one house nearer to us than another, our other senses will bear
out the view that it is nearer; for example, it will be reached sooner if we walk along
the road. Other people will agree that the house which looks nearer to us is nearer; the
ordinance map will take the same view . . .’ (Russell 1912: 31). This is Russell’s argument
that a real spatial relation between the houses corresponds to the nearer than relation of
which we are aware between certain sense data.
⁸ The remainder of this chapter follows paragraphs in VM, ch. 19, very closely, with
the kind permission of the MIT Press.
⁹ External negation, which operates on the sentence as a whole, is called ‘immunizing’
negation in (Millikan 1984). Horn (1989) gives a parallel analysis calling it ‘meta-
linguistic’ negation as opposed to ‘descriptive’ negation. The claim is that immunizing
or metalinguistic negation is not a semantic operator.
 Useless Content 113

equivalent to the presence of a representation showing the negative of that

fact. Absence of a belief is not a negative belief. Similarly, absence of per-
ceptual evidence leading one to form or confirm a belief is not perceptual
evidence that leads one to form or confirm the negative of that belief. If you
look again from another angle at what you took to be a square object but
fail this time to see that the object is square, or reach out with your hand
but fail to feel that the object is square, this by itself is not evidence against
the object’s being square. Perhaps the trouble is that you can no longer
see the object at all, or although you see it, you can’t make out its shape
against the light. Perhaps the trouble is that the object is not where it
appeared to be so that reaching out your hand to feel it you encounter noth-
ing at all. To gather evidence against the object being square, you must first
see or feel the object, and then you must see or feel that its shape is some
contrary of square, perhaps round or oblong. Gathering evidence for the
negative of a proposition is always gathering positive evidence, evidence for
some contrary of that proposition.
It follows that an ability to recognize contraries of a property through
the variety of their diverse manifestations and to recognize them as being
contraries, as being incompatible, is required in order to test one’s abilities
to identify subjects of judgment, and vice versa. The result is not an epi-
stemological regress or circle. But both of these abilities do have to be in
place before stability of judgment over time, over various perspectives, and
through diverse media of information transmission can emerge with regard
to any particular kind of subject matter. Both these abilities have to be in
place before steady evidence can accumulate that successful identifications
are being made outside the context of practical activity.
True, without doubt the first leg up is still practical. Many of the things
recognized as the same again for purposes of practical use turn out to cor-
respond to pretty good subjects or predicates for theoretical judgment as
well. (The second leg up is public language—OCCI, ch. 6; VM, chs. 9
and 19.) But the end result is the perfection of concepts by a method that
does not rely on practical uses as its criterion of success. It does not rely
on practical uses as its criterion of success, and yet this criterion has been
selected for because its use has yielded practical results in the past. Beliefs
about things outside our light cone, then, are like desires for candies that
contain saccharine. Their content is determined biologically, but the biolo-
gical functions—logically coherent thought, obtaining sweet tastes—that
determine their content are not on the same level as any direct purposes
of the genes. Nor, in the case of coherent thought, is the content determ-
ined by its use in having helped to fulfill past psychological goals, except
in so far as avoiding contradiction might be considered a psychological
goal.
114 Ruth Millikan

REFERENCES

H, L. R. (1989), A Natural History of Negation (Chicago: University of Chica-

go Press).
H, D. L., L, R. E., and G, S. S. (2001), ‘A General Account of
Selection: Biology, Immunology and Behavior’, Behavioral and Brain Sciences,
24/2: 511–69.
K, S. A. (1982), Wittgenstein on Rules and Private Languages (Cambridge,
Mass.: Harvard University Press).
M, R. G. (1984), Language, Thought, and Other Biological Categories (Cam-
bridge, Mass.: MIT Press).
(1995), ‘A Bet with Peacocke’, in C. Macdonald and G. Macdonald (eds.),
Philosophy of Psychology: Debates on Psychological Explanation (Oxford: Black-
well), 285–92.
(2000), On Clear and Confused Ideas (Cambridge: Cambridge University
Press).
(2004), Varieties of Meaning (Cambridge, Mass.: MIT Press).
P, C. (1992), A Study of Concepts (Cambridge, Mass.: MIT Press).
P, C. (2000), Functions in Mind (Oxford: Oxford University Press).
R, B. (1912), The Problems of Philosophy (New York: Holt).
 6
On Thinking of Kinds:
A Neuroscientific Perspective
Dan Ryder

1. KINDS AND PSYCHOSEMANTICS — NOT A MATCH

MADE IN HEAVEN

Reductive, naturalistic psychosemantic theories do not have a good track

record when it comes to accommodating the representation of kinds. In this
chapter, I will suggest a particular teleosemantic strategy to solve this prob-
lem, grounded in the neurocomputational details of the cerebral cortex. It is
a strategy with some parallels to one that Ruth Millikan has suggested, but
to which insufficient attention has been paid. This lack of attention is per-
haps due to a lack of appreciation for the severity of the problem, so I begin
by explaining why the situation is indeed a dire one.
One of the main tasks for a naturalistic psychosemantic theory is to
describe how the extensions of mental representations are determined.
(Such a theory may also attempt to account for other aspects of the
‘meaning’ of mental representations, if there are any.) Some mental
representations, e.g. the concept of water, denote kinds (I shall be assuming
this is non-negotiable). How is this possible? Unfortunately, I haven’t the
space to canvass all the theories out there and show that each one fails to
accommodate the representation of kinds, but I will point out the major
types of problems that arise for the kinds of theories that, judging by the
literature, are considered viable contenders.¹ In general, the theories either

¹ For example, I ignore pure internal conceptual role theories, since they fail to
explain how a concept can even have an extension. No psychosemantics has ever been
given for ‘long-armed’ role theories (e.g. Harman 1987), and the early informational
theories (Stampe 1977; Dretske 1981) have been shown to suffer fatal problems with
disjunctivitis and the like (Fodor 1990).
116 Dan Ryder

attempt and fail to account for the representation of kinds, or they fall back
on something like an intention to refer to a kind—not exactly the most
auspicious move for a reductive theory.
There are a number of problems that prevent non-teleosemantic
theories from explaining how it is possible to represent kinds. A concept
of a kind K must exclude from its extension things that superficially
resemble K but whose underlying nature is different, e.g. a concept of
water excludes XYZ (Putnam 1975). Any psychosemantic theory that
depends exclusively upon the intrinsic properties (including dispositions)
of the representer to determine extension will thus fail to provide for
the unequivocal representation of kinds, by reason of the familiar twin
cases. This problem infects theories based on isomorphism (Cummins
1996), information (Usher 2001), and nomological covariance, including
(as Aydede 1997 demonstrates) Fodor’s asymmetric dependence theory in
its most recent guise (Fodor 1994).
Some information-based and nomological covariance theories avoid the
twin problem by adding further conditions on extension determination.
Whether or not these moves help with twins, they do not help with kinds.
For example, Prinz (2002) requires a representation’s content to be its ‘in-
cipient cause’, the thing that explains the concept’s acquisition. Such modi-
fied informational and nomological covariance theories inevitably fall victim
to a second problem for such theories, the problem of epistemically ideal
conditions. On a nomological covariance theory, surely it is the category that
exhibits the best nomological covariance with a representation that is its con-
tent. (What else could it be?) Since the covariation of a representation with
its content will always be better in ideal epistemic circumstances, a nomo-
logical covariance theory will always incorrectly dictate that the content of
a representation of kind K is rather K-in-epistemically-ideal-circumstances.
Getting the right content cannot be achieved by ruling out those
factors that are ‘merely epistemic’, e.g. by adding an ‘ideal conditions’
clause (Stampe 1977; Stalnaker 1984), since which factors are merely
epistemic will depend upon which factors are semantic (McLaughlin 1987).
If the symbol really means K-in-good-light, being in good light is not a
merely epistemic factor. (And if the symbol really means K-in-weak-light,
good light will not be ideal.) The only remaining way to rule out the
epistemic factors from the content of a representation on a nomological
covariance theory would be by ad hoc restriction of the class of candidate
contents to those that fit with how we in fact carve up the world. Because
we find these categories intuitive, this move can easily go unnoticed.
But this is to ignore the fact that a psychosemantics will be part of an
explanation for why those categories are the intuitive ones, i.e. why our
representations have the contents they do. The restrictions on content must
 On Thinking of Kinds 117

be predicted by a complete psychosemantics (perhaps supplemented by an

epistemology); they cannot be tacked on in order to fix a flawed theory.
An appropriate restriction might appear to emerge naturally from nom-
ological covariance theories, since on such theories, the classes of repres-
entable contents are presumably marked by having ‘nomic’ rather than
‘non-nomic’ properties in common. But making that distinction principled
while ruling out the problem contents and ruling in the legitimate contents
is a tall order (not to mention the collateral problem that arises for repres-
enting individuals). For instance, if crumpled, shirtness, and even crumpled-
shirtness are nomic properties (Fodor 1991), why on earth isn’t gold-in-
good-light? (Thus, in Dennett’s Philosophical Lexicon: ‘jerry-mander, v. To
tailor one’s metaphysics so as to produce results convenient for the philo-
sophy of mind’; Dennett 1987.) On the other hand, if we were to restrict
representable contents to scientific kinds, the content of much of our men-
tal life would go unaccounted for.
One might expect teleosemantic theories to do better, but in fact they do
not. Consider indicator teleosemantics, for instance (Dretske 1986, 1988;
Matthen 1988). If an indicator I is selected for indicating kind K, that
means I’s indicating kind K has been causally relevant to the presence of
I in this (species of ) organism, or causally relevant to I’s recruitment to per-
form some biologically relevant task. The problem that arises is a version
of Fodor’s (1990) indeterminacy complaint (which is itself a version of the
disjunction problem): why suppose that I has been selected to indicate K as
opposed to a disjunction of local signs of K? If K has been causally relevant
to I’s selection, then the local signs of K that mediate I’s ability to indicate K
have also been causally relevant to I’s selection. There is no reason to pick K
rather than local signs of K as I’s content, and, again, an ad hoc restriction
isn’t kosher. (In correspondence, Dretske has agreed that, for these reasons,
he faces difficulties in accounting for the representation of kinds and indeed
anything distal.)
A similar problem infects the basic version of Millikan’s ‘mapping’ the-
ory, which involves the selectional recruitment of ‘representation produ-
cers’ to produce representations that map onto the environment according
to a certain rule. This selectional recruitment operates via these repres-
entations’ consumers; it is the mapping of the representations onto their
contents that has, historically, explained the proper performance of one or
more of the consumers’ functions—in Millikan’s terminology, this map-
ping is a Normal condition on proper performance of the consumers’ func-
tions. She also requires the mapping be something that the producer can
bring about or effect, e.g. via a detection mechanism.
Now, whenever Millikan is tempted to suppose that it is a kind that a
representation is supposed to map onto, we may counter as follows: that
118 Dan Ryder

kind will have a cluster of typical properties, some subset of which will be
those that explain its detection (or however the mapping is brought about)
and the proper performance of the consumer’s relevant functions. Why not
say that the representation denotes either the subset or disjunction of selec-
tionally explanatory properties, rather than the kind itself? A kind is not
identical to a subset or disjunction of its properties. This seems to raise the
spectre of twin problems for Millikan—the selectionally explanatory prop-
erties could be clarity, liquidity, and potability, properties that H2 O and
XYZ share. (We shall see later that some additional resources allow Millikan
to deal with this problem, if not for the ubiquitous fly-snapping frog, then
at least for us.)
As far as I know, all other reductive theories face a relative of one of
the problems above. We must face the possibility that a reductive natur-
alistic psychosemantics cannot explain the representation of kinds directly.
However, there are also a number of non-reductive strategies for explaining
the possibility of representing kinds, strategies that appeal in some way to
the representation wielder’s intentions. Such an appeal might be used in a
reductionist project if the representations (normally concepts) mobilized in
these intentions do not themselves include any representations of kinds. I
think this hope is forlorn, but the reductionist has a lesson to learn from the
attempt.
What sort of intention would do the trick? Well, an intention to pick out
a kind, of course. For instance, the representation of specific kinds could
be accounted for by mental description. Just as one might say that unicorns
are picked out in thought descriptionally (‘a horse with a horn’—otherwise
a puzzling case, especially for the naturalist, since unicorns do not exist),
perhaps one could pick out a specific kind with ‘a φ that is a kind’, where
‘φ’ denotes some complex property whose representation can be accom-
modated by your favourite reductive theory. It would remain to give some
reductive account of the concept of kindhood; however, if concepts of par-
ticular kinds are difficult to account for with current reductive theories
of intentionality, then the concept of kindhood seems even more difficult.
Further, how would such a concept be acquired without prior concepts of
specific kinds as examples?
Perhaps a further non-reductive move could be made, filling out the
concept of kindhood descriptionally. A well-developed account of kind-
hood exists that could be put to use in this way. According to the ‘uni-
fied property cluster’ account (Boyd 1991; Kornblith 1993; Millikan 1999,
2000), a natural kind is characterized by a set of correlated properties,
where some further principle explains why they are correlated, and thus
why reliable inductive generalizations can be made over them. For example,
water is a substance with multiple correlated properties like liquidity in
 On Thinking of Kinds 119

certain conditions, clarity, the ability to dissolve certain other substances,

etc., where these ‘surface properties’ are explained by water’s nature or hid-
den essence, namely its molecular structure. This pattern of regularity
organized around a ‘source of correlation’ (as I call it) is not restricted to
chemical natural kinds. In the case of biological kinds, these correlations are
due, not to an underlying chemical structure, but to the common history
shared by their members.
Millikan (1998, 1999, 2000; see also Bloom 2000) extends the uni-
fied property cluster account beyond natural kinds. Non-natural (but real)
kinds also have multiple correlated properties unified by some explanatory
reason. Artifacts, for instance, will often have correlated properties because
they serve some specific function (e.g. screwdrivers), because they originate
from the same plan (e.g. Apple’s iMac), or because they have been copied
for sociological reasons (e.g. a coat of arms and its variants). Even kinds
of events and processes are sources of correlation, for instance Hallowe’en,
biological growth, and atomic fusion. (Millikan also points out that indi-
viduals fall into the same pattern, although we will not be concerned with
them here.)
Now the non-reductionist account of a kind concept becomes ‘a φ that
is a member of a class sharing a syndrome of properties with a common
underlying explanation’ or something like that, mentioning in φ some spe-
cific features of the syndrome (and perhaps restricting it to ‘around here’).
This seems to be the idea behind the theory of ‘psychological essential-
ism’ (Keil 1989; Medin and Ortony 1989), and at least one philosopher has
appealed to it in order to explain the possibility of representing kinds (Prinz
2002). Again, though, this seems cold comfort to the reductionist, for at
least two reasons. First, the sophistication of this conception is such that
the potential for reduction is not getting any clearer (not to mention the
fact that it may put kind concepts beyond the reach of a sizeable propor-
tion of the population, and I don’t just mean children). Second, it seems
to betray an almost classical empiricist faith that the concepts featured in a
kind’s syndrome can ultimately be understood (presumably through many
steps of analysis), not as kind concepts themselves, but as some other sort
of concept. (Perhaps as representations of properties transduced by percep-
tual systems? This is the extreme view that Prinz’s theory entails, although
he does not assert it.) This seems unlikely, especially for biological kinds.
For example, maleness is a biological kind characterized by a biological syn-
drome including, for instance, produces sperm, which is its own biological
kind, etc.
Doubtless there is much more to be said about the descriptionist strategy.
For instance, the ‘division of linguistic labour’ (Putnam 1975) is available to
the descriptionist, whereby people can supposedly think of kinds through
120 Dan Ryder

language, by deference to experts on the referents of public language terms.

(But how do the experts think of kinds? And does this deference not involve
concepts of kinds, e.g. the kind word ?) Or the description may be ‘natur-
alized’ by giving it a causal role reading (but how can causal roles determ-
ine extensions?) I will not pursue this line of inquiry any further here; for
one thing, there are many independent problems that accrue to descriptional
accounts of concepts generally (Fodor 1998; Millikan 2000). A direct reduc-
tionist approach would clearly be welcome.
Teleosemantics is the reductionist approach I advocate, and the teleo-
semanticist can learn from the non-reductive appeal to further intentions.
The standard teleosemantic strategy is to take what might appear to be
intentions, an agent’s purposes, and reconstrue them as functions, or bio-
logical purposes. This, I suggest, is what we should do with psychological
essentialism, eliminating the need for a complex, reduction-resistant con-
ception of kindhood on the part of the agent. What we need is for the
mind’s representational system, or a part of it, to have the function of indic-
ating kinds (Dretskean formulation), or that it be supposed to map onto
kinds (Millikanian formulation), or something similar. That is, kindhood
itself, as characterized by the unified property cluster account, needs to be
selectionally relevant to the representational system.
This, in fact, is a way of understanding what Millikan has attempted
over the course of a number of publications (1984, pt. ; 1998, 2000),
except that she claims selectional relevance not for kindhood, but some-
thing broader that includes kindhood as a subtype (she calls it substance-
hood). The account I present below was developed independently, but
has much in common with Millikan’s. What is particularly interesting is
that my story comes from the neuroscience, whereas hers is derived from
abstract psychological considerations. This convergence is surely a sign of
truth! The main idea underlying my proposal is that the brain’s predict-
ive network was selected for because of the way it interacts specifically with
kinds (actually, the more general class of ‘sources of correlation’). Its spe-
cial predictive capacities are dependent upon its interacting with kinds qua
kinds, so kindhood, as characterized in the unified property cluster account,
was selectionally relevant to the design of our representational system. One
function of our representational system, therefore, must be characterized
with reference to kindhood, and this function ultimately underlies our rep-
resentation of kinds. But getting to this conclusion will require considerable
stage-setting. First I will present a particular teleosemantic framework, and
apply it to representation acquisition (for concepts of kinds are acquired );
then I will outline the neuroscientific details that yield an answer to our
main question.
 On Thinking of Kinds 121

2. THE TELEOSEMANTICS OF MODELS

Many representations may be understood in teleosemantic terms. Although

a tire gauge carries information about pressure, temperature, and volume, it
represents only pressure because that is what it is supposed to carry inform-
ation about (Dretske 1986, 1988). A map of Oconomowoc, Wisconsin,
is two-dimensionally similar to both Oconomowoc and Blow-me-down,
Newfoundland, but it only represents Oconomowoc because that is the loc-
ation to which it is supposed to be two-dimensionally similar. (And a sheet
of paper, physically indistinguishable from a map of Oconomowoc, that
is not a map but rather a section of wallpaper is not supposed to be two-
dimensionally similar to anything, and as a result, does not represent any-
thing.) Bar graphs, pictures, and many other representations can be treated
analogously. For all of these representations, there is some type of relation
that they tend to enter into with the things they represent, and which thing
they represent appears to be the thing with which they are supposed to be so
related, or to which they have the function of being so related.²
Rather than taking indicators (Dretske 1988, 1995), or pictures, or words
to be the analogue of mental representation, I believe that neuroscience and
psychology recommend that we adopt the representational paradigm of mod-
els. Some of the most familiar examples of models are scale models, like a
child’s model airplane, or a model of a building that is to be constructed.
Models capitalize on isomorphism. Isomorphism is a relation between two
structures (e.g. spatial structures), where a mapping of elements from one
structure to the other preserves some pattern of relations across the map-
ping.³ This mirroring of a pattern of relations is what makes a model use-
ful. When our access to the thing a model represents is somehow restricted,
we can use the model to reason about that thing (Swoyer 1991; Cummins
1996). For instance, if we do not know what the left wing of the Spirit of St
Louis looks like, we can just consult our model to find out. That is an example

² Teleological theories typically put this ‘supposed to’ in terms of function. Here, this
stretches the normal use of ‘function’ a little; normally we say that something has the
function of doing something, not of being a certain way. However, it is convenient to use
the term to cover both sorts of supposed-tos, and this is how I shall use it. What matters
is the normativity, not the functionality per se.
³ Consider a structure S1 , where the elements of S1 are interrelated by a single type of
two-place relation, R1 , according to some particular pattern. That is, R1 obtains between
certain specific pairs of elements of S1 . S1 is isomorphic to another structure, S2 , if there
is a relation R2 (also two-place) and a one-to-one function mapping the elements of S1
onto the elements of S2 such that: for all x and y belonging to S1 , xR1 y if and only
if f(x)R2 f(y). This definition may be extended to n-place relations in the obvious way
(Russell 1927: 249–50; Anderson 1995).
122 Dan Ryder

of using the model to fill in missing information about the world (‘predict-
ive use’, broadly speaking). Another important use of models is in practical
reasoning, in figuring out how to act (‘directive use’). For instance, the scale
model of a building might be used as a guide for its construction. (Elsewhere,
I have argued that the occurrent attitudes are the causal role equivalents of
these two uses; 2002.)
Just like representation in indicators and maps, representation in models
is a functional property—mere isomorphism is insufficient. A rocky out-
crop that just happens to be isomorphic to the Spirit of St Louis does not
represent the Spirit of St Louis because the isomorphism in question is not
a normative one—the rock is not supposed to be isomorphic to the Spirit of
St Louis. A model represents because it has the function of mirroring or
being isomorphic to some other structure.⁴
Structures are composed of elements that enter into relations. When two
structures are isomorphic, an element of one is said to correspond to a partic-
ular element in the other, within the context of that isomorphism. These
two relations, isomorphism and correspondence, are promoted to being
representational properties when they become normative or functional. A
model represents a structure S when it has the function of being isomorph-
ic to S, and the model’s elements then represent the elements of S because
they have the function of corresponding to them. Thus representation in
models comes in two related varieties, one for the model, and the other for
its elements. A model of the Spirit of St Louis models the Spirit of St Louis,
while the left wingtip of the model stands in for the left wingtip of the Spirit
of St Louis.

3. MODEL-BUILDING

Any teleological theory of mental representation faces a problem if it relies

solely upon natural selection to endow content-determining functions upon
brain states. The problem is this: most of our mental representations, in-
cluding our representations of kinds, are not acquired through evolution,
but rather through learning. Suppose that there are models in the brain.
The internal model that you have of the rules of chess is not a model whose
isomorphism functions were determined by natural selection! Yet in order
for it to be a model of the rules of chess, it must have the function of mirror-
ing the rules of chess. Whence this function?

⁴ Actually, normative isomorphism isn’t sufficient for representation—the represent-

ing structure must also be actually or normatively put to one or more characteristic uses,
e.g. predictive and/or directive use. See Ryder (2002).
 On Thinking of Kinds 123

One teleosemantic strategy for answering this question is to treat

learning as a selectional process itself, leading to the acquisition of
functions (Papineau 1987, 1993; Dretske 1988). There are two types of
selectional processes that might be co-opted to perform the design role.
At the neurophysiological level, there is ‘neural Darwinism’, according to
which neurons die off in a way that is supposed to be analogous to natural
selection (Changeux 1985; Edelman 1987). However, even supposing the
analogy is good enough to support a notion of function, the empirical
evidence suggests that a selectional account of brain development and
learning is at best radically incomplete (Quartz and Sejnowski 1997).
The other possibility, at the psychological level this time, is some sort of
reinforcement learning story—‘design’ through reward and punishment.
(Both Dretske and Papineau rely on this.) The problem here is that new
representational capacities can be acquired merely through observation,
independently of reinforcement (Sagi and Tanne 1994; Bloom 2000).
In this section, I describe a framework that yields an account of non-
selectional learning in the form of model acquisition. In the next section,
I apply this framework to the brain. Briefly, my story is that evolution
designs a model-making machine, and the operation of this machine con-
stitutes learning. (The general framework is closely related to and probably
an instance of Millikan’s account of ‘derived relational proper functions’.
For the purposes of my argument here, though, I prefer to leave it open
whether Millikan’s more general story is correct.)
I will start with the intentional design of models, like the model airplane.
When someone produces a model of the Spirit of St Louis, they typically
consult the actual plane in producing the model. (Of course, they may do
so indirectly by consulting photographs, for example.) When we consider
the model produced, and ask the question ‘What is this a model of ?’, one
way of answering our question would be to tell us what object was used as a
template for the model. Since the Spirit of St Louis was the template for the
model produced, it is a model of the Spirit of St Louis, and not some other
plane that it happens to be isomorphic to.
Note how this already begins to move us away from a dependence upon
intentional design, which is necessary if we eventually want to apply the
account to learning. Suppose the model designer has an intention to pro-
duce a model of the Wright biplane, but (mistakenly) uses the Spirit of
St Louis as his template. Is the model that gets produced a model of the
Wright biplane, or the Spirit of St Louis? There are considerations on both
sides. We can further reduce the involvement of intentions by moving to
a case of automated model production. Consider the following device, ‘the
automatic scale modeller’, designed to produce static models. It takes some
object as input, and produces a mould from the object. Next it shrinks
124 Dan Ryder

the mould. Then it injects a substance that hardens inside the mould, and
finally it breaks the mould and ejects a small-scale model of the original
object.
Why is it that we can say that the scale model this machine produces is
a model of the original object? Suppose the original object is the Spirit of
St Louis. (It is a big machine!) There need not be any intention to produce
a model of the Spirit of St Louis at work here. Perhaps someone just set
this model-making machine loose on the world, letting it wander about,
making models of whatever it happens to come across. (Of course, there
were intentions operative in the production of the machine; what we have
eliminated is any specific intention to produce a model of the Spirit of St
Louis.) The scale model produced is a model of the Spirit of St Louis simply
because the plane is what served as a template for production of the model.
The function of this machine is not to produce isomorphs of particular
things; it has the more general function of producing isomorphs of whatever
it is given as input. Each individual model inherits its function of mirroring
some specific object O from this general function, and the fact that O is
the input that figures in its causal history. Consequently, for any particular
model the machine produces, we must know that model’s causal history in
order to know what it represents.
But there is something else we need to know: the machine’s design prin-
ciples. In our example, the spatial structure of the model represents the spa-
tial structure of the thing modelled. But the model has a number of other
structural features besides its spatial structure; for example it has a density
structure. However, these other structural features are not representation-
al. Even if it fortuitously turned out that our scale model of the Spirit of St
Louis has exactly the same density structure as the Spirit of St Louis, the dens-
ity structure of the model would not correctly represent the density structure
of the plane (just as a black-and-white TV doesn’t correctly represent the col-
our of a zebra). This is because if the scale model happened to have a density
structure that mirrored the density structure of the real plane, it would be
entirely by accident, in the sense that it would not be by design.
A model-making machine is designed so that certain specific types of
relational features of input objects will cause the production of a specific
type of isomorphic structure. Those features of the input object that, by
design, determine the isomorphism for the automatic scale modeller are spa-
tial relations—and so spatial relations are the only relations the model rep-
resents, that it has the function of mirroring. Similarly, the only relational
features of the model that are structured by the input object, by design, are
spatial relations. Thus the design principles of the automatic scale modeller
tell us that only the spatial features of the model it produces do any rep-
resenting. When supplemented with the production history of a particular
 On Thinking of Kinds 125

model, the design principles can tell us exactly what that model and its ele-
ments represent, i.e. what the model has the function of being isomorphic
to, and what its elements have the function of corresponding to in the con-
text of that isomorphism. Similarly for any other model-making machine:
the machine’s design principles plus the causal history of a particular model
will tell us what that model represents.
Note that the automatic scale modeller is capable of producing inaccur-
ate models. Perhaps a piece of the machine falls off during its operation,
and introduces a lump into the model of the plane. This model says some-
thing false about the plane’s structure. Alternatively, it may be that the
general design principles for the machine fail in certain unforeseen circum-
stances, e.g. perhaps deep holes in an object cannot be fully penetrated by
the modelling clay. In both of these types of inaccuracies, the machine fails
to produce what it is supposed to produce, namely a structure spatially iso-
morphic to its input.
In the automatic scale modeller, there are two stages to the production of
a genuine model with a specific content. I propose that we can apply these
two stages of model production to the brain, in particular to the cerebral
cortex (because the thalamocortical system is the most likely brain structure
to subserve mentality). The first stage is the design of the model-making
machine, either intentional design (the automatic scale modeller) or evol-
utionary design (the cortex). The second stage is exactly the same in both:
template-based production of specific models according to the design prin-
ciples of the machine, as determined by the first stage. This is what it is to
acquire new representations through (non-reinforcement) learning.
If we suppose that the seat of the mind, the cerebral cortex, is designed
(by natural selection) to build models of the environment, the crucial ques-
tion that arises is this: what are the design principles of the cortex? In the
next section I will describe, from a functional point of view, the essentials
of these design principles according to the SINBAD theory. First, though,
a little preview of how this foray into neuroscience will help us eventually
answer the question we started out with, of how it is possible to represent
kinds.
The type of models the cortex is designed to build are dynamic models.⁵
The elements of a static model and the isomorphic structure it represents
are constants, like the position of the tip of the plane’s wing, and the pos-
ition of the tip of the model’s wing (relative to other points internal to the
plane). By contrast, in a dynamic model the elements in the isomorphic

⁵ The earliest extended physicalist discussion of the dynamic isomorphism idea, and a
defence of its relation to the mind, occurs in Kenneth Craik’s The Nature of Explanation
(1943). See also Cummins (1989) and McGinn (1989, ch. 3).
126 Dan Ryder

structures are variables. Rather than mirroring spatial structure, a dyna-

mic model mirrors covariational structure. For instance, a model used for
weather prediction might have elements that correspond to positions in the
atmosphere, where these elements can take on different values depending
upon whether there is likely to be rain, snow, a hurricane, or clear sky at
that position. The values of the elements in the model covary in complex
ways, and those covariation relations are meant to mirror covariation rela-
tions in the atmosphere. (Weather models are used only predictively, to
fill in missing information (about the future); but if we had the ability to
manipulate some weather-affecting variables, such models could be given
directive use as well.)
The SINBAD theory is a theory of cell tuning. A neuron ‘tunes’ to an
entity x in the environment when it adjusts its connections from other
neurons such that it has a strong response to x and a weak response to
other items (see Figure 6.1). The important thing to note is that cell tun-
ing occurs under the influence of the environment. I think that we ought to
conceive of multiple cells’ tuning as a process of template-based production
of dynamic models.
It was important that the automatic scale modeller was designed so that
the represented structure influenced the production of the representing struc-
ture in the model. What that means for an automatic dynamic modeller
is that the regularity or covariational structure of the environment must
influence the structuring of the model. A simple example of such dynam-
ic structuring under the influence of the environment would be classical
learning by association (Figure 6.2a). The associationist supposes that we
begin with internal items that are already ‘tuned’ to particular things in the
environment. Taking the neurophysiological point of view, suppose that
the internal items are neurons, and that one neuron begins its life tuned to
Cell response

Stimulus dimension

Figure 6.1. Cell tuning.

 On Thinking of Kinds 127

(a)

(b)

External variable

Internal variable

Figure 6.2. In (a), mirroring of pairwise correlations (associationism); in (b), mir-

roring of multiple correlations.

flashes, and another begins its life tuned to booms. Through a process of
association, the pairwise correlation between flashes and booms (in thun-
derstorms) comes to be reflected in a mirroring covariation between the
neurons tuned to flashes and booms.
There are a number of reasons why the cortical design principles can-
not be those of classical associationism. One particularly serious problem
with the associationist proposal is that it is too impoverished to explain
our capacity to reason (Fodor 1983). In any case, there is neurophysiolo-
gical evidence that the regularity structure in the environment that guides
production of cortical models is not simple pairwise correlational struc-
ture, as the associationist supposes. Rather, the template regularity pattern
is of multiple correlations, i.e. multiple features that are all mutually correl-
ated (Figure 6.2b) (Favorov and Ryder 2004). This proposal also receives
support from psychology. While people tend to be quite poor at learning
pairwise correlations, unless the correlated features are highly salient and
the correlation is perfect or near-perfect ( Jennings et al. 1982), when mul-
tiple mutual correlations are present in a data-set, people suddenly become
128 Dan Ryder

highly sensitive to covariational structure (Billman and Heit 1988; Billman

1996; Billman and Knutson 1996).
Already, a special relationship between the cortex and kinds is intim-
ated. Recall the unified property cluster account of kinds; it said that a kind
is characterized by a set of correlated properties, i.e. multiple mutual cor-
relations, where some further principle explains why they are correlated.
According to the SINBAD theory, the principal cells of the cerebral cortex
are built to take advantage of this general pattern of regularity, the pattern
due to sources of correlation (including kinds). Interacting with sources of
correlation allows SINBAD networks to become dynamically isomorphic
to the environment, making them useful for prediction (and direction). My
eventual claim will be this: SINBAD networks (and thus the cortical net-
work) are designed to produce isomorphs to regularity structures involving
kinds (and other sources of correlation) specifically. The SINBAD design
principles designate kinds as part of the proper template for the cortical
model-building machine; thus we can say the cortex, and the enclosing
agent, genuinely represent kinds.

4. THE SINBAD THEORY OF THE CEREBRAL CORTEX

The relevant cortical design principles apply in the first instance to pyr-
amidal cells (see Figure 6.3), the most common neuron type in the cere-
bral cortex (70 to 80 per cent of the neurons in the cortex fall into this
class—see Abeles 1991; Douglas and Martin 1998). Like any other neur-
on, a pyramidal cell receives inputs on its dendrites, which are the elaborate
tree-like structures as depicted on the cell in Figure 6.3. A cortical pyram-
idal cell typically receives thousands of connections from other neurons,
some of which are excitatory, which increase activity, and others of which
are inhibitory, which decrease activity. (Activity is a generic term for a sig-
nal level.) Each principal dendrite—an entire tree-like structure attached
to the cell body—produces an activity determined by all of the excitatory
and inhibitory inputs that it receives. This activity is that dendrite’s output,
which it passes onto the cell body. The output of the whole cell (which it
delivers elsewhere via its axon) is determined in turn by the outputs of its
principal dendrites.
The input–output profile of a dendrite, and thus its contribution to the
whole cell’s output, can be modified by adjusting the strengths of its syn-
aptic connections, and possibly by modifying other properties of the dend-
rite as well, like its shape (Woolley 1999; McAllister 2000). An important
question in neuroscience is: what principles underlie the adjustments a cell
makes in order to settle on some input to output causal profile? Why do
 On Thinking of Kinds 129

a
principal
dendrite

cell body
(soma)

axon

Figure 6.3. A typical cortical pyramidal cell. The dendrites form the input region
of the cell, which transmits its output via the axon. There is a total of five principal
dendrites visible on this cell. (‘Dendrite’ can refer either to a principal dendrite or
to a sub-branch of a principal dendrite.) Axons from other neurons synapse on one
or more of the thousands of tiny spines covering the dendrites; inhibitory synapses
may also occur between spines.

certain connections become highly influential, while others get ignored or

even dropped? And what determines the nature of the influence they come
to exert? For short: What is the pyramidal cell ‘learning rule’? The SINBAD
theory provides one plausible answer.⁶
The proposal is this: that each principal dendrite will adjust its connec-
tions so that it will tend to contribute the same amount of activity to the
cell’s output as the other principal dendrites on the cell. So if there are five
principal dendrites, like on the cell in Figure 6.3, they will each tend to
adjust their connections over time so that they will consistently contrib-
ute one-fifth of the cell’s total output. I’ll put this by saying, ‘They try to
match each other’s activities.’ They are not literally trying, of course; it is
merely a brute causal tendency that they have. (The acronym ‘SINBAD’

⁶ For full details of the SINBAD theory, please see Ryder and Favorov (2001), Ryder
(2004), and Favorov and Ryder (2004).
130 Dan Ryder

stands for a Set of INteracting Backpropagating Dendrites, which refers to

the mechanism by which the dendrites try to match each other’s activities.)
For simplicity, consider a SINBAD cell that has only two principal dend-
rites. They are trying to contribute an equal amount, 50 per cent, to the
cell’s output; that is they are trying to match each other’s activities. And
they are trying to do that consistently, no matter what inputs they happen
to get. Suppose the cell’s two principal dendrites are connected to the same
detector, or sensory receptor. In this situation, it will be very easy for them
to match. If they both just pass their input on to the cell body without
manipulating it in any way, they will always match.
However, dendrites do not get the same inputs, as a rule (Favorov and
Kelly 1996). Thus in the typical situation, the two dendrites’ matching task
will not be trivial. Suppose, for instance, that they receive two completely
unrelated inputs. To use a fanciful example, suppose dendrite A receives an
input from a green ball detector, while dendrite B receives an input from a
whistle detector. Suppose both detectors go off at the same time; i.e. there
is a green ball present, and also a whistle sounds. So both dendrites become
active at the same level, let’s say 40 units, and they pass that activity on
to the cell body, which will become active at 80 units. The dendrites have
both passed the same amount of activity on to the cell body, so according to
the SINBAD connection adjustment principle, they will not change their
connections at all. The next time either one receives its input, it will treat it
in the same fashion as it did this time.
But remember that it was a coincidence that there was a green ball and
a whistle present at the same time. Next time, perhaps there is just a green
ball. The output of the cell will then be 40 units, where dendrite A accounts
for 100 per cent of this output, while dendrite B accounts for 0 per cent.
The dendrites have radically failed to match. The adjustment principle dic-
tates that dendrite A weaken its connection to the green ball detector, and
that dendrite B strengthen any active connections (of which there are none,
we are supposing). But it is a hopeless case; the two dendrites will never
consistently match activities no matter how they adjust their connection
strengths, because they are receiving two utterly unrelated inputs. The only
way they can match is if their inputs are in some way mutually predictable.
The most basic form of mutual predictability is simple pairwise cor-
relation. If green balls and whistles were consistently correlated, then the
two dendrites would be able to match their activities consistently. So, for
instance, if dendrite A also received a connection from a beak detector,
and dendrite B received a connection from a feather detector, the dendrites
could learn to match. The beak and feather connections would strengthen,
while the green ball and whistle connections would weaken to nothing. The
learning rule would make dendrite A come to respond strongly to beaks,
 On Thinking of Kinds 131

and dendrite B to feathers. Because beaks and feathers are consistently cor-
related in the environment, the dendrites will consistently match.
Of course, there are more complex forms of mutual predictability than
simple correlation. Real dendrites can receive thousands of inputs, and they
are capable of integrating these inputs in complex ways. So the dendrites
can find not just simple correlations between beaks and feathers, but also
what I call ‘complex correlations’ between functions of multiple inputs.
Consider another cell. Suppose that amongst the detectors its first dend-
rite is connected to is a bird detector and a George Washington detector,
and for its second dendrite, a roundness detector and a silveriness detect-
or. (Clearly detectors that no well-equipped organism should be without!)
There is no consistent simple correlation between any two of these, but
there is a consistent complex correlation—bird XOR George Washington
is correlated with round AND silvery. So in order to match consistently, the
dendrites will have to adjust their input–output profiles to satisfy two truth
tables. The first dendrite will learn to contribute 50 per cent when [bird
XOR George Washington] is satisfied, and the second one will learn to con-
tribute 50 per cent only when [round AND silvery] is satisfied; otherwise
they will both be inactive (output = 0). Since these two functions are cor-
related in the environment, the two dendrites will now always match their
activities, and adjustment in this cell will cease.
Consistent environmental correlations are not accidental: there is virtu-
ally always a reason behind the correlations. For example, the correlation
between beaks and feathers in the first example isn’t accidental—they are
correlated because there exists a natural kind, birds, whose historical nature
(an evolutionary lineage) explains why they tend to have both beaks and
feathers. What will happen to a cell that has one dendrite that comes to
respond to beaks, while the other comes to respond to feathers? The cell
will respond to birds—the thing that explains the correlations in its inputs.
Similarly, the second cell will come to respond to the kind that explains the
complex correlations in its inputs, namely American quarters.
SINBAD cells thus have a strong tendency to tune to sources of correl-
ation. Different cells will tune to different sources of correlation, depend-
ing upon what inputs they receive. Each cell’s tuning is to be explained
by a particular source, and the correlations that source is responsible for
(Figure 6.4). When this tuning process takes place over an entire network,
the network is transformed so that its flows of activation come to mir-
ror regular variation in its containing organism’s environment. Where the
environment has some important variable—a source of correlation—the
network will have a cell that has tuned to that source of correlation. And
where there is a predictive relation among sources of correlation, the net-
work will be disposed to mirror that relation. The activities of the network’s
132 Dan Ryder

External variable

Source of correlation

Sensory input

SINBAD cell tuned to source of correlation

Figure 6.4. A SINBAD cell tunes to a source of correlation by selecting sensory

inputs from the mutually correlated external variables that constitute that source of
correlation’s syndrome of features.

cells will covary in just the way that their correspondents in the environ-
ment do. In short, the network becomes dynamically isomorphic to the
environment.
The reason that a cortical SINBAD network develops into a dynamic iso-
morphism is that cells’ inputs are not only sensory, but also (in fact primar-
ily) derived from within the cortical network. A cell’s tuning is guided, in
part, by these intracortical connections (Phillips and Singer 1997). Tun-
ing—changing a cell’s dispositions to react to the environment—occurs
through the modification of a cell’s dispositions to react to activity in oth-
er cells, mediated by intracortical connections. It is these latter dispositions
that come to mirror environmental regularities.
Remember, on the classical associationist picture, a pairwise correlation
in the environment comes to be mirrored in the brain (Figure 6.2a). On
the SINBAD picture, it’s not simple pairwise correlation that comes to
be mirrored in the brain, but patterns of multiple correlations, plus their
sources (Figure 6.4). In Figure 6.5, where one SINBAD cell’s inputs come
from other SINBAD cells in the cortical network, it can be seen that,
through the process of cell tuning, the regularities obtaining among the
 On Thinking of Kinds 133

cell receiving input from other SINBAD cells

Figure 6.5. When a SINBAD cell receives intracortical inputs from other SIN-
BAD cells, the relations among sources of correlation come to be mirrored in those
connections. (The mirroring is shown as bidirectional since cortical connections
tend to be reciprocal, though the two directions are mediated by distinct ‘cables’).

sources of correlation upon which these cells depend for dendritic matching
will come to be reflected in their intracortical connections. Since sources
of correlation are interrelated both within and across levels (cats are related
to fur and to mice, water is related to taps and salt, and grass is related
to greenness and to suburbia), an extensive network develops, as a cell’s
dendrites come to use other cells’ outputs in finding a function that allows
them to match.⁷ A cell may start with a tenuous correlational seed,⁸ but
this subtle sign of those correlations’ source is enough to put the cell on a
path towards discovering the multitude of regularities in which that source
participates. As the cell achieves more and more robust dendritic matching,

⁷ Naturally, in getting their dendrites to match, cells can take advantage not only of
intrinsic features, but also of relational features of sources of correlation.
⁸ If there is no correlation available, the cell’s activity will be low, and it will elaborate
its dendrites in search of new inputs until it is able to find a correlation (for a review of
dendritic growth, which occurs throughout life, see Quartz and Sejnowski 1997). If it is
still unsuccessful, at some point the cell will ‘give up’ and degenerate (Edelman 1987).
134 Dan Ryder

the correlational seed ends up producing a complex dendritic tree, which

realizes complex functions relating that source of correlation to many others
(or rather cells that have tuned to them).
Thus in tuning to a source of correlation, the dendrites of a particular
pyramidal cell find mathematical functions that relate that source of correl-
ation, not only to sensory inputs, but also to other sources of correlation
via intracortical connections. In this way, the connections among cells gain
characteristics that dispose flows of activity to mirror regularities involving
these sources.⁹ This is extremely useful. When a SINBAD cell is activated,
this amounts to the network ‘inferring’ the presence of a particular source
of correlation, both directly from sensory input, and indirectly from other
cells that are active owing to the presence of the source of correlation to
which they have tuned. A cell that has tuned to a particular variable has
a large number of sources from which it can obtain information about
that variable, from numerous sensory input channels and also neighbour-
ing cells. (This is due to the inductive richness of sources of correlation,
and the capacity of a cell’s dendrites to make use of much of this richness
in learning to match.) If one of those sources of information is blocked,
e.g. sensory inputs, the others will compensate.¹⁰ In the context of the net-
works’ dynamic isomorphism to the environment, a cell that corresponds
to the kind tiger (because that is what it has tuned to) will light up when
all that is seen is a twitching tail, or even a footprint. That is, intracortical
connections allow the network to perform the trick of ‘filling in missing
information’.
Here then, in brief summary, is how SINBAD networks operate. The
multiple dendrites on a SINBAD cell must find mathematical functions of
their inputs that are correlated. Assuming these correlations are not acci-
dental, the cell will tune to their source. In tuning to a source of correl-
ation, a cell will provide neighbouring cells with a useful input, i.e. an
input that helps their dendrites to find correlated functions. Thus these
neighbouring cells, in turn, tune to sources of correlation, and the process
repeats. The end result of this complex multiple-participant balancing act is
that a SINBAD network, richly endowed with internal links, comes to be

⁹ If you are worried that there are far too many sources of correlation our brains
need to have some cells tune to, consider the fact that in the densely interconnected
human cerebral cortex, there are somewhere between 11 and 25 billion pyramidal
cells (Pakkenberg and Gundersen 1997). Compare this to a good adult vocabulary of
50,000 words. (There is also a mechanism to prevent too many cells from tuning to the
same source of correlation—see Favorov and Ryder 2004.)
¹⁰ Of course, this will create a mismatch between dendrites; if a previously correlated
input is consistently absent, the dendrite will learn to ignore it in order to achieve a match
again with the other dendrites on the cell.
 On Thinking of Kinds 135

dynamically isomorphic to the environment from which it receives inputs.

This dynamic isomorphism mirrors the deep structure of the environment,
with elements that correspond not only to sensory features, but also to the
kinds, natural and otherwise, and other sources of correlation around which
environmental regularities are structured.

5. A TELEONEUROSEMANTICS FOR THE

REPRESENTATION OF KINDS

Can the SINBAD theory explain how the representation of kinds is pos-
sible? It should be uncontroversial that the cortical network, if it is a SIN-
BAD network, is a model-building machine. Clearly the cortical network is
supposed to structure itself isomorphically with regularities in the environ-
ment; the utility of this isomorphism is undeniable, for filling in missing
information about the world, and in practical reasoning. But our main
question, of how it is possible to represent kinds, turns upon the nature
of the specific design principles of a SINBAD network. The design prin-
ciples of a model-making machine dictate its general function, and thus
what type of structure it represents. We have seen that SINBAD networks
have a strong tendency to become dynamic isomorphisms that mirror regu-
larities organized around sources of correlation. The result we now want to
get is that this tendency is teleofunctional: that the cortical SINBAD net-
work was designed to develop such isomorphisms, and consequently that
SINBAD cells are supposed to correspond to sources of correlation. Given
the analysis of model representation from Section 2, it would follow that
SINBAD cells represent sources of correlation. Since kinds form one type
of source of correlation, we would have shown how it is possible to repres-
ent kinds.
Note the contrast between this approach and the descriptionist’s. The
descriptionist begins with a (complex) representation of a cluster of observ-
able properties that typically characterize some kind. But, observes the de-
scriptionist, no such representation can ever represent a kind—it will never
have the right extension to do so (owing to twin problems, for example).
The only way to mentally represent a kind, they continue, is to represent
it as a kind in a very strong sense: one must have a detailed conception of
kindhood, and somehow link this with the representation of the cluster of
observable properties. (In Section 1, I tried to show that this was not a very
promising approach.)
According to my proposal, this detailed conception of kindhood is not
necessary. Contra the descriptionist, it is possible for a representation that
does not include anything like a conception of kindhood nevertheless to
136 Dan Ryder

‘get the extension right’ in the case of a kind. The automatic scale modeller
produces representations of (relative) spatial points on objects in virtue of
having the general purpose or function of producing correspondences to spa-
tial points. It need not have anything like a conception of spatial-pointhood
in order to do this. Similarly, a representing device may produce repres-
entations of kinds in virtue of having the general purpose or function of
producing correspondences to kinds, while utterly lacking a conception of
kindhood. (In fact, my proposal is slightly different, of course. It says that
the cerebral cortex has a general purpose or function of producing corres-
pondences to the broader class of sources of correlation, but that still means
it can ‘get the extensions right’ in the case of kinds, since kinds are simply a
variety of source of correlation.) Does this count as representing a kind ‘as
a kind’? I don’t know. Some would take a detailed conception of kindhood
to be necessary for that phrase properly to apply. All I care about at the
moment is solving the problem described in Section 1, which was the prob-
lem of ‘getting the extensions right’, something that no other extant theory
can manage. (Perhaps SINBAD neurosemantics can also help explain how
it is possible to acquire the concept of kindhood, as well as the conception
that typically accompanies it, but I make no claims about that here.)
So all we need to show is that SINBAD cells have the general purpose
or function of corresponding to sources of correlation (or, if you prefer,
that the cortex has the function of ‘producing’ cells exhibiting such cor-
respondences). Since evolution is the designer here, we need to make it
plausible that the SINBAD mechanism was selected for the properties of its
interaction specifically with sources of correlation, that its being structured
by sources of correlation in particular confers some benefit compared to
other types of model-building (e.g. pairwise association). This is eminently
plausible. We saw that the clustering of numerous (possibly complex) prop-
erties around a source of correlation allows a cell that tunes to that source to
have multiple lines of ‘evidence’ for its presence. The result is an extremely
powerful predictive network, with multipotent capabilities for filling in.¹¹
Importantly, SINBAD cells must tune to reliable sources of multiple correl-
ations in order for the network to exhibit this sort of power; the particular
advantage of the network depends entirely upon the inductive richness of
sources of correlation. SINBAD cells are plausibly built (by evolution) to
take advantage of this inductive richness—they have a strong tendency to
tune to sources of correlation, and this tendency is what ultimately pro-
duces a rich isomorphism.

¹¹ Note that several functions may overlap on a single dendrite, and typically cells will
operate in population units, with all members of one population corresponding to the
same source of correlation. So the capacity for filling in is astronomical (Ryder 2002).
 On Thinking of Kinds 137

There are several related aspects to the way in which SINBAD cells
take advantage of the inductive richness of sources of correlation.¹² First,
this richness permits a SINBAD network, given the nature of its units, to
develop a correspondingly rich inductive network, not of scattered pairwise
correlations (as in an associative network), but of interrelated regularities
grounded in the deep structure of the environment. This richness ensures
robust prediction through redundancy —there are many ways to predict the
same thing.
Second, the inductive richness of sources of correlation facilitates future
learning. Because sources of correlation are inductively rich, once a SIN-
BAD cell starts to tune to one by discovering some of the correlated proper-
ties it exhibits (the correlational seed), the cell (given its special properties)
is in a uniquely advantageous position to discover further correlation.¹³
(This will be the case as long as its dendrites have not come to match their
activities perfectly, which they almost never will, owing to the presence of
noise in the cortical network.) We saw that in this way, a cell continually
adds to its lines of ‘evidence’ for the presence of the source of correlation to
which it is tuning, indefinitely enriching the model’s isomorphism.
Relatedly, the success of SINBAD networks depends upon a cell, dur-
ing the course of learning, receiving useful inputs from other cells in the
network. These other inputs will be much more useful to aid dendritic
matching if these other cells have tuned to real kinds, so that their outputs
carry information about real kinds. That is because environmental regular-
ities are fundamentally determined by interactions among real kinds (and
other sources of correlation). So, given the nature of the SINBAD mechan-
ism, it’s vital that cells tune to sources of correlation in order to develop a
nice isomorphism.
A way to think of it at the network level is this: the SINBAD cortical
network was selected for mirroring, not just regularities, but grounded reg-
ularities (grounded in sources of correlation). But, one might ask, how can
mirroring grounded regularities be selected for? In order for a mechanism to
be designed to mirror grounded regularities, it must incorporate some way
of picking them out. But how could any mechanism do this?

¹² They are relatives of functions that Millikan attributes to empirical concepts (in
the previous chapter in this volume, and her 2000, ch. 3): ‘accumulating information’
about substances (including real kinds), and ‘applying information previously gathered’
about substances.
¹³ This has a corollary: learning becomes much more efficient when a SINBAD
network creates correspondences to sources of correlation. Compared to the acquisition
of a model lacking such correspondences, fewer relations involving fewer variables must
be learned in order for the model to be inferentially complete. (See Favorov and Ryder
2004; Kursun and Favorov 2004).
138 Dan Ryder

Well, the SINBAD mechanism does this, as compared to its compet-

itors (i.e. other broadly ‘associative’ mechanisms, including those actually
found in other parts of the brain, like the hippocampus, basal ganglia, amy-
gdala, and cerebellum). It does not, of course, incorporate some infallible
groundedness detector, but function does not require infallibility. SINBAD
networks exhibit a peculiar sensitivity to multiple mutual correlations, and
multiple correlations are a sign of groundedness. (Related ideas commonly
surface in epistemology and metaphysics. In epistemology: corroborating
evidence indicates truth. In metaphysics: multiple causal powers that are
only contingently co-instantiated indicate a real object.) SINBAD networks
then, as opposed to other correlation-based mechanisms, have a (fallible)
way of winnowing out the useful, grounded correlations. Thus they could
have been selected for this ability, and plausibly were. So they plausibly
have the specific function of mirroring grounded regularities, in particu-
lar (given the reasons adduced above) regularities grounded in sources of
correlation.
So just as the automatic scale modeller has a dispositional ‘fit’ for pro-
ducing spatial isomorphs (and not density isomorphs), a SINBAD net-
work has a dispositional ‘fit’ for producing isomorphs to regularities centred
around sources of correlation (and not to more generic sorts of regularities).
It is precisely this fit with sources of correlation that gives SINBAD net-
works an advantage over other types of mechanism with respect to richness
and redundancy of prediction. So it is reasonable to infer that the cortical
network, if it is a SINBAD network, was (in part) designed by evolution
to come to mirror regularities specifically involving sources of correlation.
This is one of the design principles of the cortical model-making machine.
In particular, SINBAD cells were designed to tune to and come to corres-
pond to sources of correlation in the context of the SINBAD network’s iso-
morphism. Since SINBAD cells have the general function of corresponding
to sources of correlation, they represent or stand in for sources of correlation
in the models the cortex produces. And since kinds are sources of correla-
tion, SINBAD cells can represent kinds, in the sense that they can ‘get the
extensions right’.
Particular extensions are determined by these general design principles
in conjunction with the history of production of a particular SINBAD
model.¹⁴ In order to figure out which specific regularity structure is rep-
resented by a specific cortical model, we need to be able to identify what
regularity structure served as a template for that model. At a finer grain,
we can figure out what sources of correlation served as the template for

¹⁴ Here follows a brief account. For more details, see Ryder (2002, 2004).
 On Thinking of Kinds 139

production of each of the elements of the model.¹⁵ The elements of a SIN-

BAD model are particular cells that have tuned or are in the process of
tuning to a source of correlation. However, the cell’s template is not just
any source of correlation that has helped cause it to fire at some point in its
past. Something serves as a template for model production only relative to
the design principles of the model-building machine.
Call the type of structure that a particular model-making machine is sup-
posed to mirror (according to its design principles) ‘the structural type that
is proper’ for that machine. Something can both (1) be an element of a struc-
tural type that is proper for a particular machine, and (2) causally affect the
structure of a model produced by that machine, while nevertheless failing
to be a template for that model. Thus when a rock falls in through the out-
put door in the automatic scale modeller causing a lump on a model of the
Spirit of St Louis, that lump does not represent the spatial structure of the
rock. Rather, the machine has produced an imperfect model of the Spirit of
St Louis. The design principles of the machine dictate what object is being
modelled; only objects that enter through the input door are modelled.
A better example might be a machine that is designed to fly around, find
single objects (by detecting coherent outlines, perhaps), circle around while
photographing them, and compile the collection of 2D information into
a 3D model, discarding information about surrounding objects. Suppose
a rock standing in front of the Spirit of St Louis obscures a portion of the
fuselage from all the machine’s camera positions, so that it must interpolate
part of its 3D model, an interpolation that fails to mirror the actual spa-
tial structure of the plane. Given the design principles of the machine, the
correct way to describe this is as an inaccuracy in the model of the Spirit of
St Louis. Even though this portion of the model was in part caused by the
rock, it does not in any way represent the rock. (The reason that this is a
better example is because this machine’s design process is more open to the
environment; it is thus more prone to error, just as a SINBAD network is.)
SINBAD cells are designed to come to correspond to sources of correlation
through their dendrites learning to match, where this learning is dependent
upon some particular source of correlation. This is how the process is sup-
posed to proceed: a cell, which starts off with randomly weighted connections
to other cells, is exposed to a source of correlation many times. Upon each
exposure, it improves its dendritic matching owing to correlations in some

¹⁵ Regularity-structure-based template determination is more holistic and coherence-

related, while element-based template determination is more atomistic and correspond-
ence-related. Both can be relevant to content determination; where conflicts arise, the
result may be a specific kind of equivocation (see below). See also Millikan’s contribution
to this volume (previous chapter), on the relation between coherence and correspondence.
140 Dan Ryder

of the properties that have helped cause it to fire.¹⁶ These properties are cor-
related owing to their being properties of this particular source, so the cell
finally comes to tune to and correspond to that source of correlation. This
is the functionally normal route for a SINBAD model to adopt a particular
configuration, the way it was designed to work: some specific source of cor-
relation causes (or explains) each cell’s achievement of dendritic matching.
Only in this way can a cell participate in a reliable predictive network. It is
equivalent to the automatic scale modeller taking in an object through its
input door, producing a nice mould, and spitting out a perfect model.
In a SINBAD model, deviations from the way structuring is supposed to
proceed by design will be deviations from the way tuning is sup-
posed to proceed by design. These will include causal interactions with
things that have inhibited a cell from achieving its current level of matching
success, and in most cases, these inhibitors will not be templates for the cell.
For example, consider the following history of SINBAD model produc-
tion. Suppose a cell had been gradually tuning to cats. Perhaps a dog caused
the cell to fire at some point, because in certain conditions, dogs look like
cats. Let us say that the dog made three of the cell’s dendrites match, while
there was a failure to match for two dendrites. The SINBAD learning rule
made some of these dendrites modify their connections. But this does not
improve their overall matching success. The dendrites will tend to move
away from functions that pick out cats (functions they had previously been
tending towards), without taking them any closer to functions that pick out
dogs, or anything else. Without consistent ‘training’ through exposure to
multiple dogs, the dendrites are unlikely to modify their behaviour so as to
increase their sensitivity to features characteristic of dogs. (In fact, the only
thing they might improve their sensitivity towards, in this case, is this par-
ticular dog in this particular circumstance—and that is not even a source of
correlation.) Subsequent exposures to cats bring the dendrites back towards
the function that picks out cats, and the cell back towards better match-
ing success (and thus predictive utility). That response to a dog inhibited
the cell from achieving its current level of matching success; it led it away
from finding the correlated functions due to cats, the isomorphism it has
now settled on. The dog was something that affected the model, but not

¹⁶ Despite the fact that the causal relation between the activity in a SINBAD cell’s
dendrite and a source of correlation is mediated by some intervening physiology, it is
still causation in virtue of some determinate property of the stimulus. Which property
was causally relevant to the activity in the synapse can be identified by counterfactuals.
Suppose an instance of a particular shade of red causes a cell to fire. Had the stimulus
been a different shade of red, the cell would have fired anyway. However, had the
stimulus been blue, it would not have fired. Then the property that was causally relevant
to the synaptic activity was redness, not the particular shade of red, nor colouredness.
 On Thinking of Kinds 141

according to design. It features in the history of the cell’s tuning, but it did
not cause or explain its matching success, i.e. the aspect of model structuring
that occurs by design. That dog was a stray rock in the SINBAD mechan-
ism, while the kind cat was this cell’s template. A cell’s template is a source
of correlation that explains its current matching success.
This does not mean that the model-building machine was broken when it
changed so as to reduce its predictive utility; it was just functioning sub-
optimally. Also note that our conclusion that this cell represents cats is
consistent with an alternative history in which the cell permanently veers
off its course, eventually tuning to and representing dogs. In this case,
the cell would have different properties with a different explanation for its
matching success, and it would be in a different model, with a different
history—indeed, at some point in its progress, the kind cat may cease to
provide any explanation for its matching success. Its matching success may
depend entirely upon its previous exposure to dogs.¹⁷ On the other hand, if
the kind cat (as well as dog) continues to explain its matching success, the
cell will be disjunctive or ‘equivocal’, where two kinds are confused as being
the same (Ryder 2002). (See Millikan 2000 on equivocal concepts, which
certainly exist and so ought to be psychosemantically explicable.) This is
another way model design can proceed sub-optimally. It is sub-optimal
since it will lead to inductive errors.
So the design principles of the cortical model-making machine pick out,
as a cell’s template, only the things that have helped that cell achieve its
current matching success (where that is measured holding its response pro-
file and current broad environment fixed). Anything else does not explain
the creation of an internal model according to the cortical design principles.
Therefore, a single SINBAD cell¹⁸ has the function of corresponding only
to the source of correlation that actually helped it achieve the degree of
dendritic matching it has attained thus far. That is the source of correlation

¹⁷ Note that this avoids a problem that arose for Dretske’s (now abandoned) account
of representation in Knowledge and the Flow of Information (Dretske 1981). In this
book, Dretske identifies the content of an indicator with the information that was
instrumental in causing it to develop a particular sensitivity during its ‘learning period’.
When it responds to something else after the learning period, it misrepresents. Suppose
an indicator has been responding to As, and then it responds to a B. On Dretske’s
old theory, if this latter response is part of the learning period, Bs will be part of the
indicator’s content, but if it is part of the ‘use’ period, then the indicator misrepresents
the B as an A. Which is it? As Loewer (1987) points out, the problem for Dretske is
that there is no principled way to distinguish between the ‘learning’ period and the ‘use’
period. In SINBAD neurosemantics, there is no need to distinguish between a learning
period and a use period. You just need to ask whether B explains to some non-negligible
extent, the cell’s current matching success.
¹⁸ I note again that representations that actually have a cognitive effect will typically
involve populations of SINBAD cells.
142 Dan Ryder

the cell represents. Anything else that it responds to, has responded to, or
corresponds to in the context of some isomorphism is not part of the cell’s
representational content.
So not only can SINBAD cells have the function of corresponding to
kinds in the context of an isomorphism, the details of the SINBAD mech-
anism allow us to determine exactly which kind (or other source of cor-
relation) a particular SINBAD cell has the function of corresponding to.
An element of a model represents that which it has the function of corres-
ponding to. So if all goes well, that kind will be the unique representational
content of the cell. Since SINBAD cells are the basic elements of a SINBAD
network, we can also determine which regularity structure a whole network
has the function of being isomorphic to, and thus models. Because of their
inductive richness and SINBAD’s penchant for such richness, kinds will
tend to figure prominently in these internal models. Which, in addition to
the evidence linking SINBAD to the cortex, and the cortex to the mind, is
an important reason to suppose that mental representation, at least in us, is
SINBAD representation.

REFERENCES

A, M. (1991), Corticonics: Neural Circuits of the Cerebral Cortex (Cambridge:

Cambridge University Press).
A, C. A. (1995), ‘Isomorphism’, in J. Kim and E. Sosa (eds.), A Compan-
ion to Metaphysics (Oxford: Blackwell).
A, M. (1997), ‘Has Fodor Really Changed his Mind on Narrow Content?’,
Mind and Language, 12/3–4: 422–58.
B, D. O. (1996), ‘Structural Biases in Concept Learning: Influences from
Multiple Functions’, in D. Medin (ed.), The Psychology of Learning and Motiva-
tion (San Diego: Academic Press).
and H, E. (1988), ‘Observational Learning from Internal Feedback: A
Simulation of an Adaptive Learning Method’, Cognitive Science, 12: 587–625.
and K, J. (1996), ‘Unsupervised Concept Learning and Value Sys-
tematicity: A Complex Whole Aids Learning the Part’, Journal of Experimental
Psychology: Learning, Memory and Cognition, 22: 458–75.
B, P. (2000), How Children Learn the Meanings of Words (Cambridge, Mass.:
MIT Press).
B, R. N. (1991), ‘Realism, Anti-Foundationalism, and the Enthusiasm for Nat-
ural Kinds’, Philosophical Studies, 61: 127–48.
C, J-P. (1985), Neuronal Man: The Biology of Mind (Oxford: Oxford
University Press).
C, K. J. (1943), The Nature of Explanation (Cambridge: Cambridge University
Press).
 On Thinking of Kinds 143

C, R. (1989), Meaning and Mental Representation (Cambridge, Mass.: MIT

Press).
(1996), Representations, Targets, and Attitudes (Cambridge, Mass.: MIT Press).
D, D. (1987), The Philosophical Lexicon, <http://www.blackwellpublish-
ing.com/lexicon>.
D, R., and M, K. (1998), ‘Neocortex’, in G. M. Shepherd (ed.), The
Synaptic Organization of the Brain (Oxford: Oxford University Press).
D, F. (1981), Knowledge and the Flow of Information (Stanford, Calif.: CSLI
Publications).
(1986), ‘Misrepresentation’, in R. J. Bogdan (ed.), Belief: Form, Content and
Function (Oxford: Oxford University Press).
(1988), Explaining Behavior (Cambridge, Mass.: MIT Press).
(1995), Naturalizing the Mind (Cambridge, Mass.: MIT Press).
E, G. M. (1987), Neural Darwinism: The Theory of Neuronal Group Selec-
tion (New York: Basic Books).
F, O. V., and K, D. G. (1996), ‘Local Receptive Field Diversity within
Cortical Neuronal Populations’, in O. Franzen, R. Johansson, and L. Tereni-
us (eds.), Somesthesis and the Neurobiology of the Somatosensory Cortex (Basel:
Birkhauser).
and R, D. (2004), SINBAD: A Neocortical Mechanism for Discovering
Environmental Variables and Regularities Hidden in Sensory Input’, Biological
Cybernetics, 90: 191–202.
F, J. (1983), The Modularity of Mind (Cambridge, Mass.: MIT Press).
(1990), A Theory of Content and Other Essays (Cambridge, Mass.: MIT Press).
(1991), ‘Reply to Antony and Levine’, in B. Loewer and G. Rey (eds.), Mean-
ing in Mind: Fodor and his Critics (Oxford: Blackwell).
(1994), The Elm and the Expert (Cambridge, Mass.: MIT Press).
(1998), Concepts: Where Cognitive Science Went Wrong (Oxford: Oxford Uni-
versity Press).
H, G. (1987), ‘(Nonsolipsistic) Conceptual Role Semantics’, in E. Lepore
(ed.), Semantics of Natural Language (New York: Academic Press).
J, D. L., A, T. M., and R, L. (1982), ‘Informal Covariation
Assessment: Data-Based versus Theory-Based Judgments’, in D. Kahneman, P.
Slovic, and A. Tversky (eds.), Judgment under Uncertainty: Heuristics and Biases
(Cambridge: Cambridge University Press).
K, F. (1989), Concepts, Kinds, and Cognitive Development (Cambridge, Mass.:
MIT Press).
K, H. (1993), Inductive Inference and its Natural Ground (Cambridge,
Mass.: MIT Press).
K, O., and F, O. (2004), ‘SINBAD Automation of Scientific Dis-
covery: From Factor Analysis to Theory Synthesis’, Natural Computing, 3/2:
207–33.
L, B. (1987), ‘From Information to Intentionality, Synthese, 70: 287–317.
MA, A. K. (2000), ‘Cellular and Molecular Mechanisms of Dendrite
Growth’, Cerebral Cortex, 10: 963–73.
144 Dan Ryder

MG, C. (1989), Mental Content (Oxford: Blackwell).

ML, B. (1987), ‘What is Wrong with Correlational Psychosemantics’,
Synthese, 70: 271–86.
M, M. (1988), ‘Biological Functions and Perceptual Content’, Journal of
Philosophy, 85: 5–27.
M, D. L., and O, A. (1989), ‘Psychological Essentialism’, in S. Vos-
inadou and A. Ortony (eds.), Similarity and Analogical Reasoning (Cambridge:
Cambridge University Press).
M, R. (1984), Language, Thought, and Other Biological Categories (Cam-
bridge, Mass.: MIT Press).
(1998), ‘A Common Structure for Concepts of Individuals, Stuffs, and Real
Kinds: More Mama, More Milk, and More Mouse’, Behavioral and Brain Sci-
ences, 21/1: 55–65.
(1999), ‘Historical Kinds and the ‘‘Special Sciences’’ ’, Philosophical Studies,
95/1–2: 45–65.
(2000), On Clear and Confused Ideas (Cambridge: Cambridge University
Press).
P, B., and G, H. J. G. (1997), ‘Neocortical Neuron Num-
ber in Humans: Effect of Sex and Age’, Journal of Comparative Neurology, 384:
312–20.
P, D. (1987), Reality and Representation (Oxford: Blackwell).
(1993), Philosophical Naturalism (Oxford: Blackwell).
P, W. A., and S, W. (1997), ‘In Search of Common Foundations for
Cortical Computation’, Behavioral and Brain Sciences, 20: 657–722.
P, J. (2002), Furnishing the Mind: Concepts and their Perceptual Basis (Cam-
bridge, Mass.: MIT Press).
P, H. (1975), ‘The Meaning of Meaning’, in K. Gunderson (ed.), Language,
Mind and Knowledge (Minneapolis: University of Minnesota Press).
Q, S. R., and S, T. J. (1997), ‘The Neural Basis of Cognitive
Development: A Constructivist Manifesto’, Behavioral and Brain Sciences, 20:
537–96.
R, B. (1927), The Analysis of Matter (London: Routledge).
R, D. (2002), ‘Neurosemantics: A Theory’, Ph.D. diss., University of North
Carolina.
(2004), ‘SINBAD Neurosemantics: A Theory of Mental Representation’,
Mind and Language, 19/2: 211–40.
and F, O. V. (2001), ‘The New Associationism: A Neural Explanation
for the Predictive Powers of Cerebral Cortex’, Brain and Mind, 2/2: 161–94.
S, D., and T, D. (1994), ‘Perceptual Learning: Learning to See’, Current
Opinion in Neurobiology, 4: 195–9.
S, R. (1984), Inquiry (Cambridge, Mass.: MIT Press).
S, D. (1977), ‘Toward a Causal Theory of Linguistic Representation’, in
P. A. French, T. E. Uehling, and H. K. Wettstein (eds.), Midwest Studies in
Philosophy, ii: Studies in the Philosophy of Language (Morris: University of Min-
nesota Press).
 On Thinking of Kinds 145

S, C. (1991), ‘Structural Representation and Surrogative Reasoning’, Syn-

these, 87: 449–508.
U, M. (2001), ‘A Statistical Referential Theory of Content: Using Inform-
ation Theory to Account for Misrepresentation’, Mind and Language, 16/3:
311–34.
W, C. S. (1999), ‘Structural Plasticity of Dendrites’, in G. Stuart, N. Sprus-
ton, and M. Häusser (eds.), Dendrites (Oxford: Oxford University Press).
 7
Teleosemantics and the Consumer
Mohan Matthen

Perceptual systems are automatic sorting machines. (See Matthen 2005,

pt. .) By processing the energy or chemical signals emanating from envir-
onmental stimuli, they sort these stimuli into classes on some consistent
basis. From the perceiver’s point of view, this activity culminates in percep-
tual appearance. Stimuli that have been assigned to the same class present
the same appearance: for example, things that are assigned to the same class
by the colour vision system look the same with respect to colour. Con-
versely, things that have been assigned to different classes present different
appearances: things that have been assigned to different classes by colour
vision look different with respect to colour. Thus, things that look blue—or
the same shade of blue—have been assigned to the same class by colour
vision; their blue appearance is an internally generated sign of this. Such
classes are nested: thus, two things that have been assigned to blue may at
the same time have been assigned to different shades of blue. That they both
look blue is a consequence of the first assignment; that they look somewhat
different in colour is a consequence of the second.

1. WHERE TELEOSEMANTICS FITS IN

The line of thought expressed above epitomizes what J. J. Gibson called

an ecological approach to perception. In this way of thinking, a perceptu-
al experience is not merely —as classical empiricists like Locke would have
it—an occurrence within the mind of a perceiver from which the state of
the external world can be inferred. Rather, it is an engagement with what
the sensory system takes to be an environmental object (Matthen 2005,
pt. ), and betokens a sub-personal act of classification with regard to this
object. (I take ‘sub-personal’ to be roughly equivalent to ‘modular’ in the
sense given by Jerry Fodor 1982.)
 Teleosemantics and the Consumer 147

Now, if one adopts this ecological line of thought, it is natural to ask:

What feature do objects assigned to a single class share? what, for example,
is the feature shared by things that appear blue? Is it possible to say in an
informative way what distinction a sensory system is making when it classi-
fies an object in a certain way?
One may distinguish two ways of answering this question more or less
consonant with the ecological approach. The first concentrates on the pro-
cess leading up to the perceptual experience. Things that appear blue have
occasioned a process that culminates in the perceptual experience of them
as blue. In view of this, one might propose that blue is the feature of envir-
onmental objects that explains this commonality of perceptual process, the
feature that stands at the head of a causal process with this result. But this
proposal has an obvious flaw. There are things that appear blue but are not
really blue. These things also stand at the head of a perceptual process that
culminates in the appearance of blue. But they are not blue. Equally, there
are things that really are blue, but do not look to be so. These things do not
share in the feature in question, but are blue. So it looks as if this process
proposal cannot accommodate the evident divergence between appearing to
be of a kind and really being of that kind.
This flaw can be corrected by concentrating on the range of circum-
stances, ‘normal’ or ‘standard’ conditions as they are variously called, in
which things that appear blue are guaranteed to be blue, and stipulating
that blue is the feature of environmental objects that explains the appear-
ance as of blue in these circumstances. But one may also abandon the causal
approach altogether, in favour of what one might call a semantic approach.
At the heart of this approach is the idea that perceptual experiences are rep-
resentations of environmental objects: thus, they refer to objects, attribute
features to them, and possess in consequence, a truth-value. That some-
thing is represented as blue is compatible with its not actually being blue.
So we do not have to appeal to non-normal conditions to explain how there
can be a mismatch between sensory appearance and reality. This change of
perspective comes, however, at the cost of introducing a new and contro-
versial concept—that of representation. The natural home of this concept
is in the study of communication between agents who possess intentions
and goals. It is not immediately clear how it can be extended to states
issued by automatic sub-personal systems. Let us put this problem aside
until later.
The semantic approach to perceptual experience demands a reformula-
tion of the question asked earlier. Blue is, on this approach, the feature
attributed to an object of perception by a perceptual experience of that
148 Mohan Matthen

object: it is the property a thing appears to possess just in virtue of looking

blue. Thus:
X is blue if and only if X has the feature attributed to something by its
looking blue.
So the question can now be put in the following way: what feature is attrib-
uted to a stimulus by a perceptual experience of that stimulus as blue? Or:
what feature does a thing appear to have just in virtue of looking blue?
Now, this is not, strictly speaking, a semantic question. All that semantics
tells us about simple terms is what they refer to or denote. Further, it tells
us how to compute what complex terms refer to or denote, given what their
simple components refer to or denote. The question now being considered
is about how a simple term gets its semantic value, so to speak, or about the
real-world relation that must hold between a term and its semantic value.
This goes beyond the scope of semantics itself: it is a meta-semantic ques-
tion. Summarily: it is a question not about the meaning of something’s
looking blue, but about how meaning relations get established in the world.
The situation here is similar to that concerning the causal theory of
names. David Kaplan (1989: 574) observes that this theory—the ‘histor-
ical chain theory’, as he calls it—does not tell us what a name means. The
meaning of a name is just its referent, Kaplan insists. Rather, the historical
chain theory tells us how—by virtue of what relation—a particular object
gets to be the meaning of a name. The theory offers us a genetic explanation
of a semantic relation, and is, in this way, meta-semantic. When I use the
name ‘Aristotle’, I am simply referring to Aristotle. In order to do this, I do
not need to know how this name came to be meaning-related to Aristotle
himself. When I use the term ‘Aristotle’ to refer to Aristotle, I am displaying
implicit or explicit grasp of the conventions governing the use of established
proper names. In order to do this, I do not need to know anything about
the social, religious, or legal customs by which such names get attached to
individuals, or about the means by which these name–individual connec-
tions ground naming relations in the language. This is for a meta-semantic
theory, which is of considerable interest to theoreticians of language, but
not a part of what a user of language must know.
In exactly the same way, the semantic approach to perceptual content
should not be taken as aiming to expose how a perceptual experience came
to meaning-relate to a feature of things in the world—as above, a perceiver
does not need to know this in order to grasp the meaning of an exper-
ience of something as blue. The question posed above is meta-semantic.
It seeks a specification of the relation that must obtain between a percep-
tual experience and a sense-feature if such an experience is to denote the
feature.
 Teleosemantics and the Consumer 149

One simple theory of the meta-semantics of perceptual experience states

goes like this:
Indicator Meta-Semantic Theory. A perceptual experience E attributes
feature F to its object if states of the type E occur only when caused by a
thing with feature F.
Introducing a technical term to simplify the above cumbersome locution:
E attributes F to x if an occurrence of E indicates that x is F.
This account imports ideas of causal origin from the process approach
sketched above, and fails for exactly the same reason. It does not accom-
modate the divergence of appearance and reality. It seems clear that it is
possible to misperceive something as blue. Such a misperception attrib-
utes to its object a feature that the object does not actually possess. Yet,
if attribution was a matter of indication, this would be impossible: for an
experience that attributes the feature could not have occurred if the feature
was not really there.
A second meta-semantic theory, not subject to this difficulty, goes like
this.
Teleo-Meta-Semantic (TMS) Theory. A perceptual experience E attrib-
utes feature F to an object x if states of type E have the biological function
of getting the perceiver to respond in a functionally appropriate way to x
possessing F.
Once again, simplifying by the introduction of a technical term meant
simply to abbreviate the above:
E attributes F to x if an occurrence of E is supposed to initiate the
appropriate response to x being F.
Teleosemantic theories—for the most part, I drop the ‘meta’ in what fol-
lows—treat of perceptual experience as an intermediary between an extern-
al situation and what an organism is supposed to do in that situation. The
experience tells the organism that the object of perception (i.e. x in the
above schema) has the feature (F), and that it should respond appropriately;
it is not merely an inert indicator that this situation has occurred. Fred
Dretske puts this point in terms of a distinction between ‘natural signs’
which merely indicate, and representations, which are treated as signs for
a particular purpose. He says: ‘Putting chilled alcohol in a glass cylinder
doesn’t generate a misrepresentation unless somebody calibrates the glass,
hangs it on the wall, and calls it a thermometer’ (1988: 67). It is when it
is used as a sign of the temperature that the alcohol becomes a represent-
ation. This is the central idea of the Teleosemantic Theory. (Later on, we
150 Mohan Matthen

will essay a more comprehensive defence of the idea that perceptual exper-
iences are representations. This paragraph will suffice for now.) Clearly,
something can have a certain function but not perform it in a particular
instance. Thus, the above theory is compatible with falsity of attribution.
A virtue of TMS Theory is that it seems to make the question we are
dealing with—‘What is the feature attributed to a thing by an experience
of it as blue?’—recognizably scientific. Scientists in many areas of the life
sciences—evolutionary biology, physiology, psychology, ethology, etc.—
ask about the biological function of animal organs and the various states
and conditions of these organs. The methodology of addressing these ques-
tions is contested, but constitutes fairly familiar territory nonetheless. Thus,
the Teleosemantic Theory serves as a bridge between philosophy and cog-
nitive science.

2. ENTER THE CONSUMER

TMS Theory meets an immediate point of resistance when one tries to apply
it to cognitively sophisticated organisms. What is the functionally appro-
priate response to a given perceptually represented situation—for instance,
to the experience of something as blue? Is there such a thing? In Matthen
(1988) I argued that there is not. It seemed obvious to me then that while
sensory states in primitive organisms, or vestiges of these states in sophistic-
ated organisms (e.g. the sneeze, the startle response, etc.), might lead directly
to autonomic responses, perceptual states in human and other higher anim-
als might lead to any number of responses depending on other volitional and
cognitive states. There is no determinate way that we are supposed to react
to the presence of a blue thing much less to one that merely appears blue.
Suppose I see a nearly full glass of beer on the table. It is by no means
automatic that I will pick it up and take a sip. Is it mine or somebody else’s?
Am I drunk or sober? Is it six o’clock on a hot evening when I am look-
ing forward to my first drink of the day? Or is it six in the morning when
I have found an unfinished glass that someone left there overnight? These
alternatives suggest that the perceptual state feeds into a complicated pro-
cess of contextualization, cognitive assessment, and decision-making. This
is its function, not the initiation of some autonomic response. It seems not
just simplistic, then, but flat out wrong-headed to define perceptual content
in terms of some single response which perception is supposed to initiate.
Call this the Problem of Multiple Responses.
Responding to this Problem, I suggested that the states in lower organisms
(or in ourselves) that tie into autonomic responses were not full-fledged
perceptual states but only ‘quasi-perceptual’. The function of full-fledged
 Teleosemantics and the Consumer 151

perceptual states, such as the one illustrated above, was merely to ‘detect’
(or, using the terminology introduced above, ‘indicate’—see Dretske 1988;
Matthen 1989) certain situations, leaving it up to the perceiver to decide
what ought to be done with the information so provided. Perception is
detection, I proposed, not the initiation of an appropriate response. By so
arguing, I brought my position close to that of Dretske (1988) as quoted
above. In short, I proposed:
Weak TMS Theory. E attributes F to x if an occurrence of E is supposed
to indicate that x is F.
The normative element contained in the words ‘supposed to’ is essential
here; it distinguishes TMS Theory from the Indicator Meta-Semantic
Theory. In the weaker version of TMS Theory, there is no requirement that
a perceptual state should initiate a specific response to the situation repres-
ented, as there is in the (stronger) TMS Theory articulated earlier: indeed,
the claim was that perceptual states proper (as opposed to quasi-perceptual
states) do not initiate specific responses.
This proposal immediately raised Ruth Millikan’s ire (1989/1993). Mil-
likan is, of course, a pioneer of teleosemantic theories (though her focus
was originally on language and communication, rather than perception),
and as such she subsumes perceptual states under the broader category of
representations. A representation, she says, is an item made for use by a
‘consumer’. It accords by a certain rule with the situation it represents
(‘accords by a certain rule’ is, I think, her version of ‘supposed to indic-
ate’), and the consumer is thereby enabled to respond appropriately to that
situation. A beaver splash representing danger is an example. If a beaver
splashes its tail in the danger-representing way, and danger in fact exists,
then other beavers in the vicinity, the consumers of this representation, will
do what they are supposed to do in the presence of danger. If it splashes
its tail when there is no danger, then these consumers will thereby be given
reason to do what would have been appropriate had there been danger, but
not what is appropriate in the actual non-threatening situation. Potentially,
this could cause the beaver’s normal activities to be disrupted. Thus, the
accuracy of the representation is, as Millikan says, a normal condition of the
success of the consumer’s actions. (In response to the Problem of Multiple
Responses, this should be amended to read ‘normal condition of the success
of the consumer’s choice of actions’.)
The Problem of Multiple Responses to a given situation is beside the
point, Millikan argued. What is important is that a connection exist be-
tween a given cognitive state and the situation or feature that it is supposed
to represent. With regard to the glass of beer, the accuracy of my perception
is a normal condition for the success both of drinking it (when that is
152 Mohan Matthen

functionally appropriate) and of throwing it away (ditto)—both actions

would be guided by the same perception. The correspondence of my per-
ceptual state to the situation that it represents is thus a normal condition for
the success of any action I might take with respect to it. Thus, ‘representa-
tional content rests not on univocity of consumer function but on sameness
of normal conditions for those functions’. To summarize: the same fact-
representation can be a normal condition for a variety of different actions.
Thus, according to Millikan, there is no need to revise or weaken TMS
Theory. What is needed is to realize that fact-detection underwrites all the
different uses that might be made of a representation.
Of course, this is not to deny that representations have a function
of detecting that which they represent. Organisms have evolved in such
a way as to be able to perform certain responsive functions—functions
that demand different actions in response to different environmental
conditions. This means not only that they will be able to effect actions
in its repertoire, but that they will be able to discriminate one situation
from another, and perform the right action in the right circumstances.
This demands a detector to make sure that the effector is well informed of
the circumstances to which it is meant to respond. The function of such
a detector will be that of issuing token representations that indicate what
circumstances obtain, at least with accuracy sufficient not to disrupt the
task at hand.
Thus, I take it that Millikan did not mean to disagree with the idea that
the function of perceptual states is ‘detecting’ or ‘indicating’—issuing a
representation that properly accords with a certain state of affairs (except
in so far as it might not have been clear that ‘detecting’ a situation was
merely another word for ‘being in accord by a certain rule’ with it). What
she meant to challenge was the idea that there is no single function if there
is no single response. ‘Matthen is . . . looking pretty squarely at the repres-
entation consumers, but at what it is the representation’s job to get these
consumers to do, rather than at normal conditions for their proper opera-
tion,’ she complains. This focus on initiating a response blinded Matthen
(i.e. me) to the single function of representations, she seems to suggest.
This is wrong. I did not say that perceptual states had no single function;
on the contrary, I said that they had the single function of detecting a cer-
tain situation—of registering that the situation has occurred and putting
this information at the service of the perceiver’s effector organs. I arrived at
the same conclusion as Millikan, but in slightly different terminology by a
slightly different route. Millikan too was proposing a Weak TMS Theory in
so far as she supposes it to be the function of a representational state simply
to be in accord with a particular situation. She takes it that when we are
specifying the function of the representation, the specific action-choice for
 Teleosemantics and the Consumer 153

which the consumer uses this representational state is irrelevant. All that is
relevant is that any action-choice will be disrupted by the representation’s
not being in accord with the situation it represents.
Millikan, however, went much further than I did with detecting.
Whereas I had been content to assume that in very low-level organisms,
the representational state that accorded with environmental situations was
the very same as the one that initiated the action, Millikan—rightly, I
think—insisted on distinguishing between the detector function and the
effector-triggering function even in simple ‘pushmi-pullyu’ representations,
as she calls them (1995)—states that combine detector and effector
functions in a single package.
Imagine that a certain cell-body in a unicellular organism enters into a
particular chemical state when a particular situation obtains, and that this
chemical state initiates a certain behaviour in the organism. (Perhaps the
lack of some nutrient causes a molecule to be stripped of a nucleotide,
which in turn causes the absent nutrient to be synthesized.) The chemical
state of this cell-body both detects the occurrence of the triggering situ-
ation and initiates the appropriate response to that situation, and is, as
such, a pushmi-pullyu representation. Despite such co-location of detect-
ing and effecting, it is important to distinguish these two capacities, and to
link them in precisely the way that Millikan does, i.e. to say that the suc-
cessful performance of the detector function is a normal condition for the
successful performance of the effector function. The detector manufactures
representations; the effector is a consumer of the representation.
This is an absolutely crucial insight for teleosemantics. However, the
point goes deeper than Millikan realized. Effector functions cannot be dis-
regarded or ignored when we ask about detector functions. To be sure, it
is a function of detectors to be accurate, and this accuracy forms part of
the normal conditions of effector success. But this does not mean that we
can state the detector function in abstraction from effector function. For
we cannot properly identify which feature the detector is supposed to detect
accurately except by reference to effector function. This is what Millikan
(and I) failed to appreciate. Both of us assumed (I surmise) that we could
find the feature in question by looking ‘upstream’, i.e. by looking at the
environment. In fact, we can find it only by looking at the consumer. Or so
I shall argue.

3. WHY MULTIPLE RESPONSES ARE NO PROBLEM

Let me return to the Problem of Multiple Responses. I now want to argue

that both Millikan and I were wrong about this. We both overlooked the
154 Mohan Matthen

fact that even in cognitively sophisticated organisms, there actually is a

set of autonomic responses to perceptual experiences, and that these are
available for the identification of the feature attributed to objects by percep-
tual experience. There is hence no need to remove from TMS Theory the
reference to a functionally appropriate response (though it is important to
recognize that it implies both a detector and an effector function).
The reason why this set of autonomic responses was not evident in the
first place is that they are not bodily actions which the organism uses to act
upon its environment, but self-directed changes to what might be called its
representational or epistemic state. The difference between perception in cog-
nitively sophisticated organisms and in detector functions in pushmi-pullyu
representations is not that the reaction to the former is ‘free’ while those
to the latter are autonomic, but rather that in the more evolved organisms,
perception controls bodily action mostly through the mediacy of epistemic
operations. However, perception still evokes autonomic responses in cog-
nitively sophisticated organisms: in particular, and crucially for its func-
tion in these organisms, it feeds autonomically into a change of epistemic
state.
At a very simple level, consider the phenomenon known as habituation.
If you present a young baby with a blue stimulus repeatedly, its initial
interest and attention dies away—on the fifth trial or so, it ignores the
presentation. If after such a series of blue presentations, you present the
baby with a red stimulus, it regards this new presentation with something
approximating the interest and attention it gave the original blue item. The
series of blue representations changes the baby’s inner state, which leads it
to treat a further repetition with a lower level of interest than it would have
met at the start of the series.
Notice how sensory classification ties into this pattern. The baby reacts
with less interest to the fifth blue presentation because each stimulus in the
series was co-classified with the others. It reacts with renewed interest to the
red presentation because this one was differentiated from the previous ones.
Thus the sorting function of colour perception is essentially involved with
the reaction pattern. More abstractly, response depends on sensory similar-
ity or dissimilarity. It is apparently functionally appropriate that organisms
should react less strongly to the occurrence of a stimulus similar to some-
thing that they have recently encountered, and perceptual sorting is (among
other things) a device for detecting whether what they are encountering
now is similar to or different from what they encountered earlier. In short,
habituation is an autonomic response to perceptual sorting, and perceptu-
al sorting is a detector function for habituation. Notice that regardless of
what the baby (or some cognitively sophisticated adult) does with the series
of presentations, she cannot avoid the change of internal epistemic state.
 Teleosemantics and the Consumer 155

Imagine an actor who is told to react with pretended surprise to the blue
presentation. She rehearses this action again and again, and perfects a look
of surprise. She has not avoided habituation, but has simply ignored it in
her outward reaction. There are multiple responses to the repeated blue
presentation, then, but there is also an autonomic response.
Habituation illustrates a more general phenomenon. Any perceptual state
alters the state of inner epistemic ‘organs’, with the consequence that re-
sponses to further presentations of that or other stimuli are altered. In
computational terms, these epistemic organs have been represented, ever
since Donald Hebb, as networks of connected neurons. Every new per-
ceptual state affects the strengths of the connections between neurons in
these networks. In more cognitive terms, one might say that the organ-
ism maintains a set of inner expectations, and every incoming perceptual
state automatically occasions an update of these expectations. With respect
to these expectations, a new stimulus will reinforce an expectation to the
degree that is experienced as similar to occurrences encountered before, or
weaken an expectation to the degree that it is dissimilar to what has been
encountered before. Such mechanisms for creating, maintaining, and extin-
guishing expectations are, as Quine (1969: 306) argues, innate. For ‘There
could be no induction, no habit formation, no conditioning, without prior
dispositions on the subject’s part to treat one stimulation as more nearly
similar to a second than to a third.’ This implies that there can be no
learning without an unlearned, i.e. innate, capacity to measure perceptual
similarity.
Broadly speaking, these expectations are of two sorts. First, we have
expectations concerning individual objects—about where they will be,
what colour or shape they will be, etc. Second, we have expectations about
associations among features: about what size goes with what weight, what
colour goes with what taste, and so on. Though these expectations change
measurably only after a series of perceptual presentations, the most plaus-
ible way to represent the connection between perception, on the one hand,
and learning or memory, on the other, is to posit that each and every per-
ceptual state has some effect on some expectation or ‘epistemic’ net. (The
term ‘epistemic’ is perhaps inflated in this context. The point is that even
very simple organisms are capable of forming environmental representa-
tions that last longer than a single interaction with that environment. In
the case of habituation, each presentation lasts only a second or so, but the
change to the response disposition lasts much longer. The term ‘epistem-
ic’ is used to mark this longer duration, the change of dispositions over the
long run; no suggestion of reasoning or justification is intended.)
In view of the occurrence of such autonomic epistemic responses, we
may elaborate TMS Theory: the responses initiated by perceptual states
156 Mohan Matthen

are epistemic, especially in cognitively sophisticated animals. Briefly, the

similarity of two things depends on how similar the epistemic response to
them is. Two exactly similar things will occasion exactly the same response.
For example, if R1 and R2 are exactly similar red chips, and a baby has
been habituated to blue, then they will evoke the same reaction of attention
and surprise. On the other hand, a violet chip which is somewhat similar in
colour to both blue and red will evoke a less surprised reaction, or perhaps
evoke surprise less often. A sense feature like blue comprises items to which
the response is the same, up to some degree of similarity. (We shall discuss
these points more fully in Section 4 below.)

4. FITTING PERCEPTION TO THE CONSUMER

We need not dwell further on the actual constitution of autonomic epi-

stemic responses. What they are in any given organism is something for
psychologists to discover. The point that is important for TMS Theory
is merely that there are such autonomic responses. I want now to return
to the question posed at the beginning. When objects are assigned to a
single sensory class, what feature are they represented as possessing? What,
for instance, is the feature attributed to something by the experience of it
as blue?
I said that when two things look the same (up to some degree of
similarity), they will occasion the same epistemic response. Return to the
phenomenon of habituation. Whenever a presentation of blue occurs, the
‘surprise value’ of blue is reduced. In other words, a blue-presentation
strengthens the expectation that blue occurs in the present environment.
After a certain number of repetitions, the strength of this expectation rises
above a threshold value, and the perceiver no longer diverts its attention in
order to look at new blue-presentations. On the other hand, a presentation
of red has a different effect. Having experienced blue repeatedly, the
perceiver is surprised by the sudden appearance of red, but would not
have been surprised by another appearance of blue. These presentations
are sorted by colour, and two are presentations of the same colour if they
lead to the same response, and of different colours if they lead to different
responses.
This gives us a hint as to how we should go about answering the question
posed. We ask: what is the principle of sorting that the perceptual system
employs? What do two stimuli that evoke the same response share—in
what respect are they alike?
There is a tendency in the philosophy of perception to search for an
answer to this question in the external world. A detector is supposed to
 Teleosemantics and the Consumer 157

have states that accord with certain states of the environment. So it seems
that the right way to approach the question is to find out what states of
the environment detector states are supposed to accord with. The question
we are asking demands more than that we establish a correlation between
detector states and environmental states. What we are looking for here is a
theory of function within which to embed any such correlation. One pop-
ular theory about colour perception is that it is supposed to detect surface
spectral reflectance (cf. Hilbert 1987, 1992; Matthen 1988). The idea is
that the colour vision system sorts things by surface spectral reflectance, and
that an organism has colour vision just in case it is able to discriminate such
reflectances. This is an organism-independent way of characterizing sens-
ory feature. Having looked at the real-world functional correlate of colour
perception states, we arrive at a characterization of colours in purely phys-
ical terms. Call this the ‘upstream’ approach: it looks upstream from the
perceiver to find the meaning of perceptual representations.
The upstream approach neglects the consumer. Every detector function
is associated with effector functions within the same organism. As we saw
in the previous section, perceptual functions are associated with epistemic
functions. These epistemic functions result in new representations, which
are associated with various further effector functions, and so on. All these
functions together form a system that serve the organism in its interac-
tions with the environment. Consider the detector system from this point
of view. It will work well if it sorts things together when they are ‘sup-
posed to’ be treated the same for these epistemic purposes, and sorts them
differently, i.e. differentiates them, when they are supposed to be treated
differently. To revert to Millikan’s insight about the consumer, perceptu-
al classification is correct if it serves the organism’s effector functions; it is
incorrect if it disrupts these functions. For example, the perceptual sorting
function that serves habituation is supposed to group things together when
it is appropriate to ignore a new presentation as not constituting news. This
is the ‘downstream’ approach. It attends to the results of perceptual sorting
activities, not to the stimuli that occasion these activities, for the meaning of
perceptual experiences.
This point of view is dictated by a simple fact. Biological function
arises out of evolutionary history. The function of an organ is that feature
of it that contributed positively to the selection history of the type of
organism in which it occurs. In a system of organ functions like the one
sketched above, functions must be coordinated. If a sensory system co-
classifies things that need to be differentiated for the normal functioning of
effector organs, the organism suffers. Conversely, if the organism develops
activities that demand a different classification scheme than its detector
organs provide, it will suffer. Early teleosemantic theories concentrated on
158 Mohan Matthen

how perceptual states evolved to correspond with environmental states. It is

more to the point to ask why one set of environmental states is singled out
rather than another.
This question is answered by the following thesis:
The Coevolution Thesis. Perceptual systems coevolve with effector sys-
tems. Their function is to provide effector systems with information specific
to the performance of the behaviours produced by the effector systems.
Though it may be true that a particular colour vision system is supposed to
detect surface reflectances, why it settled on surface reflectances and how
it carves up the domain of surface reflectances depends on the epistem-
ic effector it serves. Ask the downstream question first: how must stimuli
be classified if success in epistemic effector activities is to be enabled? The
answer could well turn out to be: stimuli must be classified according to
their surface spectral reflectance. However, this answer would be subor-
dinate to the downstream question in much the same way as extension is
subordinate to intension.
The Coevolution Thesis is illustrated by a recent trend in theories of
the evolution of colour vision in primates. The central idea is that prim-
ate colour vision evolved so that certain kinds of edible vegetation would
become more conspicuous in tropical forests. (For a review, see Surridge
et al. 2003.) It is not agreed among these scientists exactly what food is
involved here: some think it is a certain kind of fruit, others that it is a
certain kind of leaf. The important point for our present purposes is that
the function of colour vision in these primates is not to maximize our
discrimination of surface reflectances, but to maximize the perceived dif-
ference between the food and the background vegetation in tropical forests
so as to make the search for food more effective. Notice that this way
of understanding primate colour vision looks beyond a broad similarity
among colour vision systems in different animal phyla, i.e. that they are
all directed towards surface reflectance. It treats this broad similarity as
merely the background to concentrating on a particular problem special
to the ancestral primate environment. Thus, it potentially explains more of
primate colour vision than the broader approach can. It explains not only
why primates detect reflectance, but also why they sharply distinguish green
from red, possibly at the cost of distinguishing among greens and among
reds. A salmonid fish would not necessarily make the same discriminations
among surface reflectances, and the coevolution thesis—the downstream
approach—tells us why.
One consequence of this evolutionary history in our present disposi-
tions could be this: we become habituated to series of colour-presentations
that differed from one another but which all fell into the general category
 Teleosemantics and the Consumer 159

of background vegetation, but be maximally surprised after such habitu-

ation by something that fell into the general colour-category of food (and
vice versa). In other words, the degree of similarity experienced depends
on this historically important ecological problem: whether things are the
colour of food or the colour of background vegetation. This not only val-
idates the theory that colour is reflectance, but explains and elaborates it.
Why and how are reflectances sorted? By the similarity relation required
for the identification of food. (It is a central part of these recent theories
that primate food itself evolved so as to take advantage of primate percep-
tual discriminations: remember that it is an advantage to a plant to have its
fruit etc. eaten, since this aids reproduction. But this is not very important
here.)
Now, given the specificity of primate evolution, it is not surprising that
different animals sort things differently, even when they are sorting by ‘col-
our’, i.e. by wavelength-sensitive discrimination. Birds are specialized to
use colour for aerial navigation; honey-bees to find pollen in flowers (not
leaves or fruit); fish to discriminate objects in water-filtered sunlight. The
emphasis on the consumer—the various systems that use detector func-
tions to serve the interests of the organisms—is necessary not to solve the
Problem of Multiple Responses, but in an informative specification of why
certain features in the real world are represented by perceptual states. All
of these organisms will sort things differently by colour. Some will sort by
wavelength-sensitive variables other than reflectance, but even when dif-
ferent organisms are sorting by reflectance, they will use different sorting
principles. (See Matthen 1999 for evidence and discussion of this.)

5. HOW WE KNOW SENSE FEATURES

The emphasis on the coevolution of the provider and the consumer of rep-
resentations—i.e. of detector and effector systems within the same organ-
ism—clears up a well-known problem for the Teleosemantic Theory. In
early versions of teleosemantics, psychological data was used to show that
sense features were physically characterized kinds. For example, David Hil-
bert (1987, 1992) and I (Matthen 1988) argued that colours were surface
spectral reflectances. (In light of the argument given in the previous section,
the conclusion ought to have been that human colours were surface spectral
reflectances grouped by a particular similarity metric.) Recently, however,
it has been pointed out (by Boghossian and Velleman 1991 and Braddon-
Mitchell and Jackson 1997) that we do not as naive perceivers know colours
under this description. After all, people in ancient times knew colours well
enough, just as well as the naive observer today knows them—and surface
160 Mohan Matthen

spectral reflectances had not even been discovered. Yet, it seems that we do
know the colours. For sensory features of this sort, to experience them is
to know them. So, it appears, the Teleosemantic Theory cannot account
for the subjective content of sense features, i.e. for how they present them-
selves to us. (Actually, it is unclear why a naive perceiver should know a
meta-semantic, as opposed to a semantic, fact about the meaning of exper-
ience: nevertheless, I will take the difficulty at face value. That is: we do
have a naive grasp of colour and TMS Theory should explain this grasp. On
the other hand, TMS Theory need not stumble on the fact that the naive
colour-perceiver need know nothing of evolutionary theory.)
Clearly, the consumer’s perspective helps with this difficulty. For instead
of trying to find the definition of sense features upstream at the head of the
process that starts with a thing out in the world, the consumer’s perspect-
ive looks downstream to the effector system for this definition. Now, one
might say that the effector system possesses one kind of ‘knowledge’ of what
a detector system’s determinations mean—it knows this in terms of its own
response. We claimed at the end of Section 3 that the proper way to define
a sense feature was in terms of sameness of response: blue is that feature
of things that brings about the same response as some paradigm—the sky,
perhaps—up to some degree of similarity. Since the epistemic response to
sensory states can be assumed to be known—tacitly, but nonetheless com-
pletely—the problem of how we know sense features seems not to arise at
all within this framework. The knowledge is instinctive and contained in
how we respond to things. (Boghossian and Velleman, who attacked ‘phys-
icalist’ theories of colour, could agree since a response-defined theory is not,
by their lights, physicalist. Braddon-Mitchell and Jackson, however, were
attacking teleosemantic theories as such, and they should find the present
argument liberating.)
In the next and final section, we examine the role of perceptual experi-
ence in making explicit the kind of tacit knowledge discussed in the present
section.

6. JUSTIFYING THE REPRESENTATIONAL FRAMEWORK

So far we have not given serious consideration to the central presupposi-

tion of the semantic approach, namely that perceptual experiences repres-
ent situations outside themselves, that they are semantic entities that refer,
describe, and are either true or false. What justifies this assumption? As
noted in Section 1, the concept of representation is naturally at home in
the arena of human communication. How can we extend it to the activities
of sub-personal systems? Here again, the consumer point of view comes to
 Teleosemantics and the Consumer 161

our aid. For as we shall now see, it aids us in constructing an important

structural analogy between acts that possess conventional meaning and per-
ceptual experiences.
We have been supposing that the sorting function of a given perceptual
system coordinates with that of an epistemic effector system. The effect-
or system has several actions within its repertoire, A1, A2, . . . , An, and
these are appropriate when some target object has feature F1, F2, . . . , Fn.
The job of the detector system is to sort objects in its field of operation
into classes corresponding to these features. For example, let’s suppose that
there are two actions available to the ‘habituation system’ (as one might
call the mechanism controlling the baby’s behaviour described in Section 3
above)—A1 is the action of attending closely to a new stimulus and A2
is the action of ignoring it. Let’s suppose that after a series of blue-presen-
tations, another blue object is presented. The system will be behaving cor-
rectly if it registers ‘Blue’, and ignores the new stimulus, or registers ‘Red’
and attends to it. Its action will disrupt what the organism is supposed to be
doing if it registers ‘Blue’ and pays the new stimulus a great deal of resource-
diverting attention, or registers ‘Red’ and ignores it. Notice that this error
might well occur because in the odd lighting conditions that prevail, or,
because the visual system has become fatigued, the new blue-presentation is
not properly classified as blue.
In order for the organism to prosper, then, two conditions must obtain:
(a) The perceptual system must accurately determine which of Blue or
Red obtains, and
(b) The effector system must reliably perform Ignore if Blue obtains and
Attend if Red obtains.
The coevolution thesis stipulates that the action-routines above were shaped
in conjunction with the sameness or difference of the new stimulus with
respect to those presented before. Thus, the question ‘To what real world
feature does blue correspond?’ is answered in the first instance by determin-
ing what the Ignore routine is supposed to accomplish. Once this is done,
one can infer what relation of similarity will optimally trigger the above
routines. This is the relation that defines the sense features in question. One
can, of course, try to determine what physical relation of similarity this cor-
responds to, but, as I argued earlier, this is a subordinate question.
An alternative way of putting this point runs as follows. The equivalence
relation that defines sense features is: the system is supposed to treat x the same
way as y for purposes of initiating effector functions. If two objects are ‘sup-
posed to’ evoke the same response—more ponderously, if the evolutionary
function of the effector system will be best served by these two objects being
treated the same way—then they belong to the same detector category. If
162 Mohan Matthen

they are supposed to be treated differently, they belong to different detector

categories. Thus, similarity is that relation to previous presentations that
merits the functionally appropriate reaction Ignore. (One assumes, on the
basis of the psychological evidence, that sameness is assessed on the basis
of sensory similarity—indeed, habituation is often used as a measure of
sensory similarity—but for present purposes, this is unimportant.) Note
here that this allows for the possibility of error. The detector system we
have been discussing may well classify a newly presented object as similar
to the ones presented before, when in fact the appropriate response would
have been to treat it as dissimilar. As Millikan urged, this would disrupt the
downstream activity.
This raises a further question. It is not enough that the detector sys-
tem should determine which of the above circumstances obtains. It is also
necessary to ensure that the effector system does the right thing. Otherwise,
we are no further ahead: the effector system would be on its own as far as
‘deciding’ what to do is concerned.
Now, in some such cases, the detector system will simply do something
that forces the appropriate operation of the effector. It may, for instance,
secrete an enzyme, or give off a molecule that gets the effector system
going. Let’s call this a coercive signal: it simply causes the requisite action to
happen. A coercive signal is what Millikan calls a pushmi-pullyu represen-
tation—which is insightful, except that (as I shall now argue) such
‘representations’ are not really semantic in character.
The phenomena cited in Section 2 in connection with the Problem of
Multiple Responses indicate a situation in which such a coercive signal
might be problematic. For, as we saw, sensory output is subjected to an
elaborate process of cognitive assessment. This suggests that the sensory
determination has to enter into a number of interacting epistemic routines.
Coercive signals have to be causally appropriate to each different effector
system. A signal that fed into a number of different systems would therefore
have to be moulded to have the appropriate effect on all of these. This is
difficult to arrange. Thus, it is unlikely that the same coercive signal could
be used for a number of different interacting systems.
What is required therefore is a non-coercive signal, one that a number
of different epistemic systems can take as an indication of what the sensory
system has determined to be the case. The detector system has, as it were,
to say ‘I have determined that such-and-such is the case’, and then leave the
effector system to handle this information as it will. (The effector system’s
action may be quite deterministic, but it is not forced by a coercive signal
issued by the sensory system.) There is a classic view of consciousness that
fits the notion of a non-coercive signal well. Cognitive scientists plausibly
 Teleosemantics and the Consumer 163

think that sensory consciousness is a monitoring device that allows the

perceiver to know about the states of her own perceptual systems. This is
just another way of saying that sensory consciousness, or perceptual experi-
ence, makes the determination of sensory systems available to the perceiver.
The news of what the sensory system has determined is simply posted in the
form of a characteristic form of experience, and the perceiver determines
how to use it in epistemic operations and rational decision-making.
David Lewis (1969, ch. 4) and Bryan Skyrms (1996, ch. 5) each have
seminal accounts of how meaning emerges in non-coercive communicat-
ive situations like this. Lewis considers a plan arrived at by a sexton and
Paul Revere. Simplifying somewhat, the sexton looks out for one of two
situations:
(R1) The redcoats set out by land.
(R2) The redcoats set out by sea.
Once the sexton has ascertained which of the two situations has occurred,
he does one of two things:
(S1) Hang one lantern in the belfry.
(S2) Hang two lanterns in the belfry.
Revere for his part looks to see how many lanterns are displayed in the
belfry. Depending on how many he sees, he performs one of two actions.
(A1) Warn the countryside that R1 has occurred.
(A2) Warn the countryside that R2 has occurred.
Now, clearly the sexton can adopt one of two action plans. They are:
(X1) If R1 then S1, and if R2 then S2.
(X2) If R1 then S2, and if R2 then S1.
Similarly, Revere can adopt one of two action plans:
(V1) If S1 has been executed, then A1, and if S2 has been executed,
then A2.
(V2) If S1 has been executed, then A2, and if S2 has been executed,
then A1.
Given that both the sexton and Revere want it to be the case that A1 if
and only if R1 and A2 if and only if R2—this is the background coordina-
tion problem—they must coordinate action plans.
We can call this the Signalling Coordination Problem. Either the sexton
should adopt X1, and Revere should adopt V1, or the sexton should adopt
X2, and Revere, V2. When one of these combinations of action plans is
164 Mohan Matthen

achieved, Lewis says, S1 and S2 are signals. In Lewis’s case, the requisite
combination is achieved by agreement between the parties. But Skyrms
(1996) shows that, under natural selection, coordinated action plans have
an ‘attractive force’ of their own, and there is no need for extrinsic acts
of agreement. As he (1996: 103) says: ‘Signaling system equilibria . . . must
emerge in the games of common interest that Lewis originally considered.’
Like Lewis, Skyrms was considering signals between organisms that have an
interest in achieving a coordinated signalling system. Here, we are consider-
ing a signal coordination problem between subsystems of a single organism.
Since the subsystems of a single organism perish or prosper according to
whether the organism does, the commonality of ‘interest’ is guaranteed. If
coordination is not achieved, the organism will be less fit, and consequently
both the detector and the effector will perish, regardless of how effective
they may be considered in themselves.
Notice that if there is to be a possibility of signalling in such a situ-
ation, the sexton has to have at least as many putative signals available
to him as there are circumstances that demand different actions on the
part of Revere. But neither cares at all which communicative action plan
is adopted, as long as the mutually desired result ensues. It follows that
there is always a choice as to which signal is associated with which circum-
stance. Thus, the association between signal and circumstance is a matter
of convention: there is always a choice among possible association schemes,
and nothing matters other than that both the sexton and Revere agree on
which signal is to be used in which set of circumstances. Earlier, I stip-
ulated that there were n features that could be attributed to a stimulus
x, and correspondingly n actions, each one appropriate when the corres-
ponding feature is detected. We now see that the sensory system would
need n non-coercive signals to inform the effector organs of which of these
features is detected. The important point to note is that, as Lewis and
Skyrms demonstrate, it does not matter which signal was associated with
which feature, so long as the actions taken by the epistemic effector sys-
tems coordinated properly with the signalling code chosen by the sensory
detector system. A coercive signal has to be chosen for its effects. If I am
going to ensure that you do the right thing by directly manipulating your
body, then I must choose my actions in such a way as to achieve this end.
With a non-coercive signal, all that matters is that my action plan coordin-
ates with yours. Thus, which signal stands for which determination of the
sensory system is contingent and historical. What matters in the case of a
non-coercive signal is that the ‘speaker’ and the ‘hearer’ should coordinate
their attitudes.
Where are we now? In Section 4, I proposed that sensory systems co-
evolve with effector systems. Here I am adding a further wrinkle to that tale.
 Teleosemantics and the Consumer 165

The claim is that under certain circumstances, the effector function is such
that, instead of direct manipulation, a non-coercive signal of the detector
state needs to be sent from detector to effector. Perceptual experience is
such a signal. The analysis of these signalling problems by Lewis and Skyrms
demonstrates that the meaning of a given experience is determined by an
arbitrary coordination scheme which emerges as a part of the coevolution
of detector and effector systems. In these circumstances, it is a matter of
convention and history which signal is associated with which circumstance
for the purposes of communication between systems. It is a convention in
the sense that (a) there is a plurality of coordination equilibria, and (b) nat-
ural selection does not ‘prefer’ one to another. This is the sense in which
we can take perceptual experience to be a representation with conventional
meaning.

REFERENCES

B, P, and V, D (1991), ‘Physicalist Theories of Color’,

Philosophical Review, 97: 67–106.
B-M, D, and J, F (1997), ‘The Teleological The-
ory of Content’, Australasian Journal of Philosophy, 75: 474–89.
D, F I. (1988), Explaining Behavior: Reasons in a World of Causes (Cam-
bridge, Mass.: Bradford Books, MIT Press).
H, D R. (1987), Color and Color Perception, CSLI Lecture Notes 9
(Stanford, Calif.: CSLI Publications).
(1992), ‘What is Color Vision?’, Philosophical Studies, 68: 351–70.
F, J A. (1982), The Modularity of Mind (Cambridge, Mass.: Bradford
Books, MIT Press).
K, D (1989), ‘Afterthoughts’, in J. Almog, J. Perry, and H. Wettstein
(eds.), Themes from Kaplan (New York: Oxford University Press).
L, D (1969), Convention: A Philosophical Study (Cambridge, Mass.: Har-
vard University Press).
M, M (1988), ‘Biological Functions and Perceptual Content’, Journal
of Philosophy, 85: 5–27.
(1989), ‘Intensionality and Perception: A Reply to Rosenberg’, Journal of
Philosophy, 86: 727–33.
(1999), ‘The Disunity of Color’, Philosophical Review, 108: 47–84.
(2005), Seeing, Doing, and Knowing: A Philosophical Theory of Sense Perception
(Oxford: Clarendon Press).
M, R G. (1989/1993), ‘Biosemantics’, Journal of Philosophy, 86:
281–97; repr. in Millikan, White Queen Psychology and Other Essays for Alice
(Cambridge, Mass.: Bradford Books, MIT Press).
(1995), ‘Pushmi-Pullyu Representations’, Philosophical Perspectives, 9:
185–200.
166 Mohan Matthen

Q, W. v. O. (1969), ‘Natural Kinds’, in Quine, Ontological Relativity and

Other Essays (New York: Columbia University Press).
S, B (1996), Evolution of the Social Contract (Cambridge: Cambridge
University Press).
S, A K., O, D, and M, N I. (2003), ‘Evol-
ution and Selection of Trichromatic Vision in Primates’, Trends in Ecology and
Evolution, 18: 198–205.
 8
Content for Cognitive Science
Karen Neander

1. INTRODUCTION

I see some newspaper blown by the wind as a cat slinking, and thus I rep-
resent the newspaper as a cat. Three things are involved: (i) a representation,
presumably some neural event, (ii) its target, in this case the newspaper,
and (iii) the content of the representation, cat slinking. In this case, the rep-
resentation misrepresents its target because there is a mismatch between its
target and its content. A philosophical theory of mental content is princip-
ally concerned with the relation between items of the first and third kind.
Such a theory tries to answer the question: in virtue of what does a mental
representation have the content it has?¹ An obvious desideratum for such a
theory of content is that it gets the contents of representations right.
It’s easy to give a theory of mental content that ascribes some content to
mental representations. There’s the Today is Tuesday Theory, for example,
which says that all of our brain states have the content Today is Tuesday.²
This allows for misrepresentation because it entails that all of our brain
states are wrong six days of the week. However, it is a terrible theory because
it gets the contents of very few mental representations right. What we want
is a theory that entails that we are thinking that today is Tuesday only if we
are thinking that today is Tuesday, and that entails that we are thinking that
the plants need watering if we are thinking that the plants need watering
instead.
While this desideratum is obvious, it is surprisingly difficult to apply.
Consider the notorious case of the frog. A normal frog will snap at anything
that’s moving and suitably small and contrastive with its background (for
short, at anything small, dark, and moving). At least as philosophers tell the

¹ As I am using the term, a mental representation need not be conscious, or part of a

conscious mental state.
² I owe this nice example to Barry Loewer.
168 Karen Neander

tale, in their natural habitat the small, dark, moving things are mostly flies
that are nutritious for frogs. There has been debate about whether this or
that theory of mental content generates suitable content in this case, and
yet—and this is one of the most frustrating things for those fresh to the
debate—opinions differ as to what the correct content is. In the case of the
frog, philosophers have variously argued that the content of its visual rep-
resentation is fly; frog food; a parcel of chemicals nutritious for frogs; something
small, dark, and moving; small, dark, moving food ; or something indetermin-
ate between these.³
How can we use simple system cases to test our theories if we cannot agree
on what the content is? What we need is some independent ground for believ-
ing one content ascription rather than another. This chapter tries to provide
such a ground. Here I argue that some candidate contents serve the purposes
of mainstream cognitive science better than others do, mainstream cognitive
science being understood as that science that uses an information-processing
approach to provide operational explanations of cognitive capacities. I claim
that some candidate contents can and some cannot play a role in such explan-
ations. This is a contentious beginning but I am content to make my con-
clusion conditional on the assumption that a theory of content should try
to meet the needs of mainstream cognitive science.
Subject to this condition, if my argument here is along the right lines,
we will have a good reason to reject standard teleological theories, such
as Ruth Millikan’s (1991), as well as some non-standard ones, such as Kim
Sterelny’s (1990) and Nicholas Agar’s (1993). On the positive side, we
will also have good reason to take another look at informational theor-
ies like those offered by Jerry Fodor (1990b), Fred Dretske (1994), Pierre
Jacob (1997), and Neander (1995, forthcoming).⁴ Note that the last three
of these are teleological theories of mental content, so this chapter should
not be construed as an argument against all teleological theories.
In what follows I switch from frogs to toads. The perceptual systems of
frogs and toads are very similar and the neuroethological literature on the two
overlaps to a great extent, but toads let me make my point a little more vividly.

³ Jerry Fodor (1990a) and Kim Sterelny (1990) say it represents its target as a fly. Ruth
Millikan (1991) says it represents it as frog food and Carolyn Price (1998, 2001) says it
represents it as a parcel of chemicals nutritious for frogs. Fodor (1990b: 106), who has
changed his mind, along with Fred Dretske (1994), who has also changed his, Pierre Jacob
(1997), and I (Neander 1995) say it represents it as something small, dark, and moving.
Nicholas Agar (1993) says it is small, dark, moving food. And Dretske (1986), David
Papineau (1998), and Daniel Dennett (1995) suggest that the content is indeterminate
between these things. This is a very incomplete list of those who have participated in this
debate, but it is representative.
⁴ This chapter is from my forthcoming book Mental Representation: The Natural and
the Normative in a Darwinian World (MIT Press).
 Content for Cognitive Science 169

I hope readers will enjoy or at least endure with patience the short excursion
into toad neuroethology that takes place in Section 2. I also hope that even
those who disagree with the implications I draw from it in Section 3 will find
the issues interesting. The smaller goal is to argue for a particular content
ascription in a particular case, but the larger goal is obviously more import-
ant. It is to illustrate the way in which content ascriptions should cohere with
neuroethological analyses of relevant cognitive capacities.
As for the relation between neuroethology and cognitive science, the two
are continuous. The relevant domain of study for frogs and toads is referred
to as ‘‘neuroethology’’ not ‘‘cognitive science’’ since the latter applies mostly
to the study of our own species. However, neuroethology does the same sort
of thing for other animals as cognitive science does for us; it studies such
things as perception, motor control, and decision making in non-human
animals, including primates. Its aims and methodological tools are also
much the same despite some obvious differences, such as in the ethical con-
straints that scientists feel themselves to be under and the fact that verbal
responses by research subjects have an important place in one and none
in the other. In both cases, the aim is to understand the normal flow and
transformation of information and its neural substrate. Sometimes expli-
citly computational models are developed in both cases (e.g. compare Marr
1982 and Cobas and Arbib 1992). Moreover, biologists believe that much
of what they have learned about the anuran (frog and toad) nervous system
applies—and many of the concepts developed in studying them are applic-
able—to a wide range of vertebrate species (Ewert et al. 1983: 414). Given
this substantial and methodological continuity, I sometimes use ‘‘cognitive
science’’ to refer to both cognitive science and neuroethology, generically.

2. TOAD NEUROETHOLOGY

Elsewhere (Neander forthcoming) I explain that there is a prima facie rea-

son to think that frogs and toads have mental representations. I argue that
if we assume, uncontroversially, that their brain states have the function of
carrying information, then they have intensional states in so far as they are
not extensional. That is, sentences describing these states guarantee neither
existential generalization nor the preservation of truth-value under the sub-
stitution of co-referring terms. This creates a presumption that mental rep-
resentations are involved, which I see no reason to override. So in my view
neither the frog nor the toad should be considered merely toy examples.
From conversations with philosophers, I have the impression that many
think of anuran ‘‘prey’’-recognition as mere transduction, but this is wrong.
The it’s-merely-transduction view might have seemed justified in the light
170 Karen Neander

of the seminal paper by Jerome Y. Lettvin and his colleagues (1959), ‘What
the Frog’s Eye Tells the Frog’s Brain’, which first sparked philosophical
interest in the frog. But this is an oversimplification even of this paper.
Lettvin et al.’s claim was that ‘‘prey’’ discrimination occurred in retinal
cells, but anuran retinal cells are more complex than mammalian ones and
the relevant process was not thought to be mere transduction (even mam-
malian retinal cells do more than mere transduction). In any case, more
recent research has undermined Lettvin et al.’s claim. It turns out that fur-
ther information processing involving mid-brain structures is required for
‘‘prey’’/‘‘predator’’/‘‘other’’ discrimination. Five decades of intensive
research further on, biologists are still trying to unravel the complexities.
Some philosophers complain that the frog example is ‘numbingly fami-
liar’ but my sympathy with this is tempered by the fact that we have main-
tained an impressive collective ignorance about the real live case. Besides,
both the frog and the toad are genuinely excellent subjects for our purposes.
As I have already remarked, the anuran brain is similar in terms of broad
principles to those of many other vertebrate species. Also, while it is cer-
tainly very complex, it is nonetheless relatively simple compared to other
vertebrate brains, and so it is easier to understand. In addition, frogs and
toads have been intensively studied—they are the amphibian equivalent of
Drosophila. As a result, neuroethologists have a more complete understand-
ing of their nervous systems than they have of most other vertebrate nervous
systems (Ewert et al. 1983: 413).

2.1. Sign-Stimuli and Prey-Capture in the Toad

So here goes. Some facts. Before we enter the brain, there are some things
that can be observed from the outside. The first is that neither frogs nor
toads feed only on flies. Lettvin et al.’s (1959) classic paper was on leopard
frogs (frogs of the Rana pipiens cluster) and adult leopard frogs are not
fussy eaters. In their natural environment they eat a variety of insects,
including sowbugs, spiders, damselflies, crickets, leafbugs, spittlebugs, and
short-horned grasshoppers. So the content ascription—flies —was never
plausible given their real diet. Adult toads are not choosy either. Different
toads have different diets but they generally eat a variety of things, such
as beetles, bugs, millipedes, slugs, and earthworms. Big toads also hunt
larger creatures, such as snakes, small birds, and even frogs. So, for toads,
no content ascription that singles out a particular prey species—worms,
say—would be suitable either. Those who fancy contents of this sort
(e.g. Sterelny 1990) need to think in terms of contents that contain a long
 Content for Cognitive Science 171

list of prey-species, or else (better) a more generic content, such as member

of one of the toad’s prey-species.
Here we will be interested in the toad’s visual representation of its prey
and in its perceptual content. One and the same representation might pos-
sess motor and/or motivational content as well as perceptual content, but
since I only look at the toad’s visual system I only discuss perceptual con-
tent here. Note also that prey catching can be triggered by tactile as well as
visual stimulation, but this is under the control of a different neural path-
way and I do not discuss it further. Normal adult toads can see stationary
objects. They don’t splat blindly into walls and tree trunks. But again this is
under the control of a different pathway.
The toad’s visually induced behavioral responses to moving objects re-
veals that it can discriminate between at least three kinds of moving objects,
which (without prejudging the outcome of this discussion) we can for con-
venience refer to as ‘‘prey’’, ‘‘predator’’, and ‘‘other’’ (to remind us not to
prejudge the outcome, I’ll use scare quotes around these terms in what fol-
lows). The typical response to these is in brief to try to catch them, avoid
them, and ignore them, respectively.
A toad’s prey-catching behavior is affected by its motivational state,
which can vary according to season, time of day, and how much it has
eaten. Sated toads stop hunting, and it is thought that escaping predators
might always have precedence over hunting. There’s also a preference-
structure for selecting a prey to hunt (or, for that matter, a predator
to avoid) if more than one prey (or predator) is detected, but this is a
complexity I shan’t pursue (Cobas and Arbib 1992 attempt to model it).
The adult capacity to distinguish between ‘‘prey’’, ‘‘predator’’, and
‘‘other’’ is innate and differs from what’s found in tadpoles, which are
vegetarian. Newly metamorphosed toads that have never before been expo-
sed to prey-like or predator-like stimuli can nonetheless perform ‘‘prey’’/
‘‘predator’’/‘‘other’’ discrimination. And they can do this even if as tadpoles
they were raised in a completely homogeneous environment. However,
accuracy (e.g. in judging distance) improves with practice and the full beha-
vioral repertoire of adults develops over several weeks or more.
Finally, toads can become habituated to repeated dummy stimuli such
as moving dots on a computer screen that reappear at the same location
or a looming cardboard square that looms in the same location one too
many times. There’s also evidence that individual experience can affect
prey selection with respect to surface features of stimuli, like dots and
stripes (Ewert and Kehl 1978). It seems that toads can learn to avoid
bees and bombardier beetles (Cott 1936; Brower and Brower 1962). Plus
172 Karen Neander

positive conditioning can affect their responses: if, for instance, the odor of
mealworms accompanies feeding for a time it can subsequently strengthen
their prey-catching response and even override what would otherwise
constitute non-prey-like (‘‘other’’) features. What follows concerns their
capacity to distinguish ‘‘prey’’, prior to any such conditioning.
David Papineau (1998: 5 n. 1) wonders if frogs and the like have a
belief–desire psychological structure, and suggests that creatures lack
determinate representational content if they lack a belief–desire psycholo-
gical structure. Whether toads have a belief–desire psychological structure
depends on how demanding the notions of belief and desire are, but it’s
worth noting in passing that toads have motivational states and therefore
states that have a desire-like direction of fit, and they also have information-
al states and therefore states that have a belief-like direction of fit. In other
words, they have states that were designed to tell them what conditions
obtain and they have states that were designed to cause certain conditions to
obtain. Toad behavior is also somewhat flexible, even aside from its modest
learning potential, in the sense that it’s modified in appropriate ways on the
basis of different informational and motivational states.
Toads respond to large looming predator-like stimuli by a range of beha-
viors that biologists describe as sidestepping, ducking, puffing up, rising up
stiff-legged, excreting toxic oils, and turning, crawling, or leaping away. In
response to a prey-like stimulus, in contrast, researchers report that the toad
displays a sequence of behavioral elements, which are said to consist typ-
ically in orienting (o) toward the stimulus, stalking or approaching it (a),
fixating or viewing the prey from front on (f ), and snapping at it (s) by
lunging and/or extending its tongue and/or snapping its jaw. There is some
flexibility in how these behavioral elements are combined. The toad might
simply snap if the prey is in front and within range, and orienting and
approaching can be left out or repeated as often as required. So we might
have sequences such as f -s, o-f -s, o-a-a-f -s, o-o-a-o-a-o-o-a-o-o-a-a-f -a-f -s,
and so on.
Usually, a toad’s response to a stimulus deemed prey-like involves an ini-
tial orienting toward it, unless the toad is already so oriented. A normal
toad does not orient toward predator-like stimuli or (prior to conditioning)
other-like stimuli. But if a toad sees a prey-like stimulus move out from
behind a barrier it first moves around the barrier, which can involve turning
away from the prey.
If the toad is placed in a glass dome and a prey-like item is rotated hori-
zontally at a constant distance around the toad the sequence is o-o-o-o-o-o-o-
o-o . . . and so on, until the toad habituates, which takes about sixty seconds.
This rotating procedure has been used in some experiments (that I describe
 Content for Cognitive Science 173

below) to gauge the extent to which a stimulus counts as prey-like for the
toad. The more turns a motivated toad makes in a thirty-second interval,
the more the stimulus is considered prey-like for the toad.
Some parts of the entire prey-capture sequence are classified as fixed-
action patterns. That’s to say that once they have begun they cannot be
modified in the light of further information. For instance, if a dummy prey
disappears after a critical point in the fixation phase, the toad still snaps
and gulps and, as the neuroethologists report, often licks its mouth in seem-
ing satisfaction. As ethologists use the term, the sign stimuli for an innate
releasing mechanism for a fixed-action pattern are those features of the
environment that release or trigger the relevant behavior. The sign stimuli
for an innate releasing mechanism can be ascertained by purely behavioral
studies through the use of dummy stimuli and the careful variation of vari-
ables, a practice that goes back to the famous studies of Konrad Lorenz and
Niko Tinbergen in the first half of the twentieth century.
Behavioral studies show that the toad distinguishes ‘‘prey’’ from ‘‘preda-
tor’’ and ‘‘other’’ by some quite specific features of a moving stimulus. The
range of the relevant dimensions and the preference curves differ from spe-
cies to species; there are hundreds of species of toad. However, the story is
much the same across the range. In a famous series of studies (summarized
in Ewert et al. 1983) Jorg-Peter Ewert and his colleagues used a variety of
dummy stimuli including cardboard cutouts with three distinct configura-
tions. These consisted of (i) rectangles of constant width and varying lengths
moved in a direction parallel to their longest axis, dubbed ‘‘worms’’, (ii) rect-
angles of constant width and varying lengths moved in a direction perpendic-
ular to their longest axis, dubbed ‘‘anti-worms’’, and (iii) squares of different
sizes, dubbed ‘‘squares’’. In brief, as shown in Figure 8.1, the ‘‘worms’’ pro-
voke prey-capture behavior, although ‘‘anti-worms’’ of the same dimensions
are ignored, and the ‘‘squares’’ produce prey-capture behavior if they are
small enough and avoidance behavior when they are larger.
As you can see, the categories ‘‘worm’’, ‘‘anti-worm’’, and ‘‘square’’ do
not quite correspond with the categories ‘‘prey’’, ‘‘other’’, and ‘‘predator’’.
However, a toad’s prey tends to be worm-like. That is, it tends to be within
certain size parameters and moving in a direction that parallels its longest
axis. Again, I use scare quotes to remind us that not all ‘‘worms’’ are real
worms and not all real worms are ‘‘worms’’. ‘‘Worms’’ can be crickets or
millipedes, or cardboard cutouts, and a real worm can be an ‘‘anti-worm’’
if, for example, it is stunned, hung by its tail, and moved perpendicular to
its longest axis.
Visual ‘‘prey’’/‘‘predator’’/‘‘other’’ discrimination is not affected by fea-
tures of the stimuli that are not captured by these dummy stimuli. For
174 Karen Neander

Behavior

40 Worm
Number of turns

s
per minute

Square s
20 s s
Anti-
worm
Worm Anti- Square
worm
0
2 4 8 16 32
Length s (degrees)

Figure 8.1. The toad’s behavioral response to worms, anti-worms, and squares,
measured in turns per minute, varies with the length s of each kind of stimulus.
After Ewert (1980)

example, changing the velocity of the stimuli makes no difference, nor

does changing the style of motion from wriggling to scuttling. A minor
qualification is that some changes, such as in the direction of contrast from
a black stimulus against a white background to a white stimulus against a
black background, can affect acuity and maximum size preference.
Neuroethologists stress that this is not a case of mere feature detection,
let alone mere transduction. The shape of the stimulus is relevant and so
is the direction of motion, but the toad’s response cannot be understood
as a response to the mere summation of these two. Two items with the
same shape can produce no response or an enthusiastic response, depend-
ing on the direction of motion, and two items with the same direction of
motion can produce no response or an enthusiastic response, depending on
the shape. The response is provoked by what is called a ‘‘configural feature’’
(or sometimes a ‘‘gestalt’’): here, motion relative to shape.
Clearly, we can learn the sign stimuli for toad prey-capture behavior
before we understand the relevant information processing or the neural
substrate.⁵ This is important because unless the sign stimuli are first
identified any attempt to develop an adequate information-processing

⁵ This creates problems for Price’s (1998, 2001) solution to the functional
indeterminacy problem(s). It conflicts with her claim that her ‘abstraction condition’
rules in favor of the content toad food (she is discussing frog food, but extending the same
reasoning to toads, we get toad food). According to Price, when we determine ‘the unique
correct description’ of the function of a device and hence, on her view, the description
that determines the contents of its representations, we must not consider the internal
design of the device or that of the collaborating systems. Where ‘d’ is the relevant device,
Price says, ‘we do not need to know how d’s fellow components make their contribution.
Nor do we need to know about the design of the device itself ’ (1998: 70). This doesn’t
favor toad food over something more like item moving in a direction that is parallel to
its longest axis because we don’t need to know how the detection device or its fellow
components work (in the sense of knowing anything about their underlying structural
 Content for Cognitive Science 175

account of the toad’s capacity or to uncover the relevant neural substrate

of that capacity will fail. Ewert et al. (1983: 415) report that some
early studies failed for that reason. An adequate model of perceptual
processing cannot be developed until we know what information is
extracted from the retina. And, as Ewert et al. (1983: 415) comment, when
neurobiologists try to uncover the neural substrate of a capacity, such as
‘‘prey’’/‘‘predator’’/‘‘other’’ discrimination in the toad, the basic strategy
is to look for neurons or clusters of neurons that have the same stimulus
preferences as the creature, in this case, the motivated toad. If researchers
lack an accurate idea of what the stimulus preferences are, they do not know
what to look for. When looking for a needle in a haystack it helps at least
to know that it’s a needle you’re looking for! We’ll have cause to return to
these points later.

2.2. Information Flow and the Neural Substrate

We turn now to the nature of the neural substrate and the information
processing it performs. How does the toad distinguish prey-like stimuli?
Although toad brains are very simple compared to ours, they are, as I
remarked before, still highly complex systems, and as a consequence the
answer remains imperfectly understood. However, even an incomplete
sketch of what’s known might suffice to show that some candidate
contents for the toad’s ‘‘prey’’-representation seem gratuitous from a
neuroethological perspective.
Of course, the processing of visual information begins with the retina,
where receptor cells (rods and cones) transduce light into neural firings.
The optic nerve, which mediates between the retinas and mid-brain struc-
tures, contains retinal ganglion cells. The receptive field of a retinal gangli-
on cell is the area of space from which light (or its absence) can affect it.
In the toad the retinal ganglion cells have receptive fields composed of two
concentric fields. Those that appear to respond differentially to ‘‘worms’’,
‘‘anti-worms’’, and ‘‘squares’’ are the R2, R3, and R4 cells, which have
an excitatory inner center and an inhibitory outer surround, meaning that
stimulation of the inner region excites the cell whereas stimulation of the
outer region inhibits it. (In some retinal ganglion cells this inhibitory and
excitatory organization is reversed.)
R2, R3, and R4 cells differ with respect to the size of their excitatory
receptive fields (ERFs), the strength of their inhibitory receptive fields
(IRFs), and the kind of stimulus that excites the center and inhibits the

or functional design) in order to know the sign stimuli that trigger the prey-catching
response. We can discover them while treating the frog or toad as a black box.
176 Karen Neander

Table 8.1. A comparison of three classes of retinal ganglion cells in a typical toad.

Neuronal class Approx. ERF diameter (deg.) IRF strength Preferred stimulus

R2 4 +++ Dimming
R3 8 ++ Dimming or brightening
R4 16 + Brightening
Source: Adapted from Ewert et al. (1983: 444).

surround. In the common toad, R2s have the smallest receptive fields,
with center diameters of 4◦ . They are primarily ‘‘off-center’’, meaning
that their centers respond best to a dimming of light. R3s have somewhat
larger receptive fields, with center diameters of 8◦ , and they have ‘‘on–off-
centers’’, meaning that their centers respond well to either dimming or
brightening. R4s have the largest receptive fields, with centers that have
diameters of 16◦ and they respond best to an increase of light. Some of
these properties are compared in Table 8.1.
Since the surround inhibits the cell, each cell responds best when the
entire center is stimulated and none of the surround is. Thus, an R2 cell will
respond most strongly when a dark circle entirely fills its center’s receptive
field. These cells can thus provide information about changes in illumin-
ation in the visual field, which can be used to extract information about
the size, shape, and motion of moving stimuli. The response patterns of
R2, R3, and R4 retinal ganglion cells to the ‘‘worm’’, ‘‘anti-worm’’, and
‘‘square’’ configurations are given in Figure 8.2.
If we compare these responses to the behavioral response patterns of the
toad, shown in Figure 8.1, we can see that none of these retinal ganglion
cells have excitation patterns that mirror the response of a motivated toad.
For instance, the behavioral response increases with the length of the worm-
like stimuli, but none of the retinal ganglion cells show the same preference.
So neurobiologists conclude that ‘‘prey’’/‘‘predator’’/‘‘other’’ discrimina-
tion requires further processing beyond that performed by the retinal gan-
glion cells.
The retinal ganglion cells primarily extend to the optic tectum (T), a
mid-brain structure, and also to the thalamic pretectum (TH), in addition
to several other neural structures. Different kinds of retinal ganglion cells
go to different layers of the optic tectum. There’s also a crossing over, with
retinal ganglion cells from the right eye crossing to the left tectum and those
from the left eye going to the right tectum. Neighborhood relations are
preserved, which means that nearby retinal ganglion cells that record from
nearby retinal receptors and thus from nearby regions of visual space project
onto nearby parts of the tectal layers. In this sense, the tectum contains
multiple maps of the visual field.
 Content for Cognitive Science 177

A. R2 cells B. R3 cells
40 40
Impulses per second

Worm
Worm Square
20 Anti- 20
worm
Square
Antiworm

0 0
2 4 8 16 32 2 4 8 16 32

Length s (degrees)

C. R4 cells
40

Square
20
Antiworm
Worm
0
2 4 8 16 32

Figure 8.2. Responses of retinal ganglion cells to three kinds of moving stimuli.
After Ewert (1980)

If the optic tectum is removed, visually induced prey-capture and preda-

tor-avoidance and indeed all responses to moving stimuli cease. Since the
avoidance of stationary barriers remains intact it is understood that the lat-
ter is under the control of a different neural pathway. The optic tectum
is involved in locating moving stimuli but it is thought to do more than
merely locate them. It appears that distinct pathways within the tectum
control the frog’s turning away from predators and turning toward prey
because these capacities dissociate, which is just to say that they can be dis-
rupted independently given sufficiently small lesions in the area (King and
Comer 1996).
In so far as the recognition of visual stimuli as prey-like can be local-
ized, neurobiologists consider the activation of a certain class of cells in the
optic tectum—the T5(2) cells—to be the best candidate. As some express
it, activity in the T5(2) cells ‘reflects a good approximation of the probabil-
ity that the stimulus fits the prey category’ (Ewert et al. 1983: 450). If you
compare the activation pattern of these cells, as shown in Figure 8.3c, with
178 Karen Neander

the response patterns of the motivated toad, as shown in Figure 8.1, you
can see that the match is quite close. It is sufficiently close for neuroetholo-
gists to think that the T5(2) cells might be the ‘prey-recognition neurons’.⁶
The response pattern of these T5(2) cells is largely explained as a balance
of the inputs from two other classes of cells, an excitatory input from anoth-
er class of tectal cells, the T5(1) cells, and an inhibitory input from some
thalamic pretectal cells, the TH3 cells. The activation patterns of TH3 and
T5(1) cells are shown in Figures 8.3a and b. Electrode implantation and
lesion experiments provide evidence that the TH3 cells have this inhibitory
effect. Electrical stimulation of the TH3 cells reduces the response of T5(2)
neurons. And if the thalamic pretectum is removed, the toad’s prey-capture
response becomes disinhibited, so that the toad acts as if everything that
moves is prey: it will then orient toward its own extremities, toward large
predator-like cardboard squares, to the hand of the experimenter, and so
on. Smaller lesions in the thalamic pretectum produce the same response
with respect to smaller parts of the visual field. Neuroethologists conceive of
this ensemble of TH3, T5(1), and T5(2) cells as something like an AND-
gate (Ewert et al. 1983: 442), or as an AND-gate with a NOT-gate on one
of the inputs (it does not have a discrete on–off character, but the ana-
logy with computational components described as AND-gates is apt). A
schematic presentation of the proposed operation of this ‘‘gate’’ is given in
Figure 8.4.
Other areas of the toad’s brain also mediate visual ‘‘prey’’/‘‘predator’’/
‘‘other’’ discrimination. For example, cells in the telencephalon inhibit the
Number of impulses

30 30 Square 30
Square Worm
20 Worm
Anti- 20 20
Square
per second

worm
Anti-
10 10 worm 10 Anti-
Worm
worm
0 0 0
2 4 8 16 32 2 4 8 16 32 2 4 8 16 32
(a) TH3 cells (b) T5(1) cells (c) T5(2) cells
Length s (degrees)

Figure 8.3. Responses of thalamic TH3 cells, tectal T5(1) cells, and tectal T5(2)
cells. Only (c), the T5(2) cells, has the same pattern as the behavioral response of
the toad, shown in Figure 8.1. After Ewert (1980)

⁶ The match is not perfect. One difference is that the maximally stimulating length
of the worm-like stimulus is 16◦ for behavior and only 8◦ for the T5(2) cells (for further
discussion of this point, see Camhi 1984: 230–7).
 Content for Cognitive Science 179

+
T5(1)
T5(2)
TH3 −
A. +
T5(1)
T5(2)
TH3
−

B.
+
T5(1)
T5(2)
TH3
−

C. +
T5(1)
T5(2)
TH3
−

D.
+
T5(1)
T5(2)
TH3
−

Figure 8.4. The T5(1), TH3, T5(2) ensemble. The response of the T5(2) cells is
primarily controlled by the inhibiting influence of the TH3 cells and the excitatory
influence of the T5(1) cells. After Ewert (1980), from Carew (2000)

activity of the thalamus. Their removal in the poor toad results in hyper-
excited visually induced escape behavior and eliminates visually induced
prey-capture behavior altogether. However, as I say, the areas that are
primarily responsible for visual ‘‘prey’’/‘‘predator’’/‘‘other’’ discrimination
are the optic tectum and the thalamus. In sum, it is thought that visually
induced prey-capture is primarily mediated by the optic tectum moderated
by the thalamus (and that visually induced predator-avoidance is primarily
mediated by the thalamus moderated by the optic tectum).

3. IDENTIFYING THE CORRECT CONTENT

It is almost time to turn to the implications of the neuroethological find-

ings for the content of the relevant representation. But first, what is the
relevant representation? Philosophers have talked of the toad’s or frog’s
perceptual representation, but what neurological state, event, or process
has been at issue? As we have seen, neither the image on the retina nor
180 Karen Neander

the firing of the optic fibers (the retinal ganglion cells) turns out to be the
‘‘prey’’-representation. Our best bet to date is that ‘‘prey’’-recognition has
only occurred once T5(2) excitation has occurred.
For this reason, I shall speak of a high frequency of action potential activ-
ity in a T5(2) cell—hereafter ‘‘+T5(2)’’ or ‘‘T5(2) excitation’’—as the
relevant representation. This is better than the usual vague or worse talk but
it could be an oversimplification. There are a number of different features of
neuronal events that could carry information. One is the average rate of fir-
ing of neurons, and this seems to be involved in this case. However, it might
not be a case of local coding (i.e. a single cell for a single feature). Instead,
a cluster of cells (e.g. with overlapping tuning curves) might be involved.
Luckily, the issue does not substantially impact the main argument in this
chapter. Only certain details would need changing.
Assuming that the relevant representation is a +T5(2), the question is
‘‘What is the content of a +T5(2)’’? Before this case can be used to test our
theories of mental content we need an independent basis for attributing one
content to it rather than another. In Sections 3.1 and 3.2, I suggest that
we need to observe the coherence constraint: if the contents of mental rep-
resentations are to play a role in explaining cognitive capacities, they must
cohere with the relevant information processing. At the risk of laboring the
point, let me stress that this is not offered as a theory of mental content.
Coherence is an intentional notion, but that’s fine and dandy here because
I am not offering a theory of mental content at this point. I am offering a
heuristic or a principle to be used in identifying contents in simple systems
so that we can determine pre-theoretically (i.e. pre-philosophical-theory-of-
content-ly) what the correct content is.

3.1. Localization Content

It is easiest to start with the localization content of +T5(2)s. This has not
been the subject of controversy in the philosophical literature. The contro-
versy has focused on what a +T5(2) represents, not on where it represents it
as being. However, in part for that very reason, the localization content can
be a useful illustration of the coherence constraint.
Recall that different retinal ganglion cells receive input from different
retinal receptor cells, which have different receptive fields, and also that the
different retinal ganglion cells extend to different cells within the relevant
layer of the optic tectum, so that different T5(2) cells also have different
receptive fields. Although the what-component of the information carried
is the same for every T5(2) cell, the where-component can differ. In a nor-
mal toad, excitation of one T5(2) cell carries the information that there’s
something (let’s just call it a whatsit for the moment) within its receptive
 Content for Cognitive Science 181

field, whereas excitation of a different T5(2) cell carries the information

that there is a whatsit within its receptive field. How precise is this localiza-
tion information?
As yet, I’ve not mentioned that the T5(2) cells are thought to be mon-
ocularly driven cells with relatively large receptive fields. If so, excitation of
each single cell of the type normally reflects, as the neuroethologists say, the
probability that a prey-like stimulus is present in a relatively large region of
the visual field. These cells do not normally carry precise information about
stimuli location on any dimension: left–right, up–down, near–far. They
are precise enough for orienting toward prey but not for accurate snapping,
and it is thought that more precise localization is carried out by different
sets of tectal and thalamic neurons. One hypothesis is that another class of
tectal cells, the T7 cells, has the function of extracting more precise inform-
ation about location from the overlapping receptive fields of T5 neurons.
Also (thalamic) TH6 and (tectal) T3 neurons, though monocularly driv-
en, are apparently sensitive to motion in the near–far dimension, so they
might provide information based on the changing size of the image on
the retina (things cast a bigger image as they approach) or from disparit-
ies in the image when the toad turns (the closer the stimulus the more its
retinal image moves). In addition, orienting brings the stimulus into the
toad’s restricted binocular range and more precise information about loca-
tions could then be derived during fixation. T1(3) neurons are a candidate
here since they are binocularly driven.
The reasoning outlined in the preceding paragraph is the kind of reason-
ing that the neuroethologists use. Indeed, it is the reasoning they use, almost
verbatim (Ewert et al. 1983: 455). It is also (more or less) the kind of reas-
oning that we should use when we look for independent (pre-philosophical-
theory-of-content) reasons for attributing contents. In line with this, I
suggest that content ascriptions should, in general, be motivated from
below by an understanding of the mechanism underlying the information
processing and from above by the information-processing explanation of
the relevant capacity. I elaborate on this in the remainder of this subsection.
Let’s look at the localization content of a particular T5(2) cell. I’ll refer to
its +T5(2)s as Rs, and to a particular token (instance) of them as r. Let L
be the receptive field of this cell, and let L− be the location of the stimulus
(inside L) that causes r to be tokened. Also let L+ be an area larger than and
encompassing both L− and L.
What is the localization content of r? Does it have the content that there
is a whatsit at L, L− , or L+ ? Rs normally reflect with some degree of prob-
ability that a prey-like stimulus is in each of these locations, so by itself this
does not discriminate between them. I suggest that the content for Rs in
general and r in particular is that there is a whatsit in L, even though r’s
182 Karen Neander

L+
L

L−

Figure 8.5. What is the localization content of r, an excitation of a T5(2) cell

caused by the worm located at L− ? Is it that the stimulus is at L− , L, or L+ ? The
receptive field of the T5(2) cell that produces r is L

stimulus is located at L− and is in L+ as well as in L, and even though r car-

ries the information that this stimulus is located at L− and is in L+ as well as
in L. Assuming you agree, the task is to figure out why this is so. Are there
identifiable principles involved?
First ask, why does it seem overly specific to say that r’s content is that
there is a whatsit at L− ? In part it is because we think that the content of r
should be the same as the contents of other Rs (i.e. other +T5(2)s produced
by the same cell). Although r carries the information that the stimulus is
at L− , this is mere happenstance, for on other occasions when this cell fires
the stimuli that provoke the Rs could be outside L− , elsewhere in L. The
mechanism that normally informs Rs has no special causal or informational
relation with L− that it does not also have with the rest of L. It has the func-
tion of informing Rs and hence r about stimuli in L, generally, as opposed
to only stimuli at L− , more specifically.
Next ask, why is it too imprecise to say that the content of r is that there
is a whatsit in L+ ? It carries the information that the stimulus is in L+ as
well as in L, and this is not mere happenstance. Normally, all Rs carry the
information that there is a whatsit in L+ as well as in L. However, Rs are
normally provoked by stimuli in L rather than elsewhere in L+ , and this is
not mere happenstance because the mechanism that normally informs Rs
has a special causal or informational relation with L that it does not have
with the rest of L+ . It has the function of informing Rs and hence r about
stimuli in L, as opposed to other areas in L+ . R-production is normally
insensitive to whatsits that are in L+ if they are not in L.
Notice as well that the wrong content ascription can interfere with its
explanatory value. If we say that the content of r is a whatsit is in L+ , we
undermine the explanation of the toad’s orienting, because it makes it a
puzzle how the toad orients as precisely as it does. (Unless we want to admit
 Content for Cognitive Science 183

that our content ascription is irrelevant to explaining the orienting.) And if

we say that the content of r is a whatsit is in L− , we undermine the explana-
tion of the toad’s orienting again, because it makes it a puzzle why the toad
does not orient more precisely than it does. It makes it a puzzle why, after
r, the toad’s brain still has to figure out the more precise location of the
stimulus before it can send snapping instructions to its motor neurons.
The coherence constraint is admittedly a bit vague but I hope that this
illustration helps. The general idea is that content ascriptions need to make
sense in the context of, and be suitable for playing a role in, an operational
explanation of the relevant cognitive capacity.

3.2. What is Represented?

Finally, we come to what is represented. Transcribing from frog to toad,
popular candidates for what +T5(2)s represent are a member of a prey spe-
cies, toad food (i.e. a packet of chemicals nutritious for toads), something mov-
ing in a direction that parallels its longest axis, and toad food that is moving
in a direction that parallels its longest axis. Which should we choose for the
purposes of mainstream cognitive science?⁷
I start with what the neuroethologists say. The truth is that they rarely
if ever utter sentences beginning with ‘the content of a +T5(2) is . . . ’,
or ‘+T5(2)s mean . . . ’, or ‘+T5(2)s represent . . . ’. However, the toad’s
‘‘prey’’/‘‘predator’’/‘‘other’’ discrimination is often used in neuroethology
texts to illustrate what is referred to as object recognition (see e.g. Camhi
1984, ch. 7; Carew 2000, ch. 4) and in this context neuroethologists do
speak in terms of (a) ‘‘prey-recognition’’, (b) ‘‘the recognition of wormlike
stimuli’’, (c) ‘‘the recognition of preylike stimuli’’, and (d ) ‘‘the recognition
of the configuration of visible features’’ (i.e. those related to the size, shape,
and motion relative to shape of the stimulus).
Different scientists do not say different things so much as use a similar
range of expressions. For example, Jeffrey Camhi says,
the toad’s T5(2) neurons are good candidates for prey-recognition neurons. (Camhi
1984: 232; emphasis added)
And then, in the next paragraph, he adds,
it is implicit in the definition of recognition neurons that they must not only
respond selectively to the particular object being recognized (in this case a moving
wormlike stimulus), but they must actually be cells that the brain uses in recogniz-
ing this object. (Camhi 1984: 233; emphasis added)

⁷ I hope it’s clear that none of this is meant to solve the so-called functional
indeterminacy problem. It’s one thing to argue for a given content, and another to show
that one’s theory delivers it.
184 Karen Neander

Thus, in the space of two paragraphs, Camhi speaks of the recognition

of prey and the recognition of a moving worm-like stimulus. Because the
italicized in (a) through (d ) are used interchangeably in this way, I suggest
that (a), (b), and (c) are all shorthand for (d ), which is the most precise but
also the most cumbersome. We have just seen (a) and (b) being used inter-
changeably, and (b) and (c), which refer to worm-like and prey-like stimuli,
respectively, presumably refer to stimuli that are worm-like or prey-like
with respect to the configuration of visible features, and are thus presum-
ably equivalent to (d ).
I have to concede, to those who dislike talk of content in the case of
such simple systems, that there is little explicit talk of error concerning what
is represented. However, there is no hesitation when it comes to talk of
error with respect to the localization content in this or similar cases (e.g.
see Carew 2000: 65–70; Grobstein et al. 1983: 334–7) and neuroetho-
logists do talk about the toad trying to catch ‘inappropriate’ stimuli. For
example, if after ablation of the thalamus the toad orients toward a large
looming square or the experimenter’s hand (remember that a normal toad
only orients toward prey-like stimuli) this might be described as a response
to an ‘inappropriate’ stimulus (e.g. Carew 2000: 115).
It is interesting that the normal toad’s response to a cardboard cutout
‘‘worm’’ is not spoken of as inappropriate (as far as I have seen). On the
contrary, it is treated as a paradigm case of the appropriate response. It is
also interesting that if a toad orients toward toad food or toward a mem-
ber of one of its prey species that is moved perpendicular to its longest axis
the toad’s response is counted as inappropriate. This kind of case is neg-
lected in the philosophical literature. Philosophers often find it intuitive
that the toad misrepresents its target if it snaps at a cardboard cutout with,
as I would say, the right configuration of visible features. But what of the
case where the toad snaps at something with, as I would say, the wrong
configuration of features?
Suppose that we kill a millipede and string it up by its tail and move it
perpendicular to its longest axis past a hungry toad. If the toad snaps at it,
is it representing correctly or incorrectly? Different candidate contents give
different answers. If the content is toad food, the relevant +T5(2) is correct.
Ditto if the content is a member of one of the toad’s prey species. But if the
content is specified in terms of the relevant configuration of visible features,
it is incorrect. Which should it be? Here I think we should side with the
neuroethologists’ talk of what is inappropriate. Not simply because that is
how they talk, but because this is what coheres with our understanding of
the information processing and the underlying mechanism.
 Content for Cognitive Science 185

Several considerations speak in favor it. One is the fact (at least prior
to any conditioning) that a normal toad does not snap at such an ‘‘anti-
worm’’. Only an abnormal toad, such as a toad with an ablated thalamus,
reliably snaps at ‘‘anti-worms’’. It seems unfortunate in a theory of content
if it implies that correct representation requires abnormality.
Putting neurological impairment aside, the contents toad food or member
of a prey species also seem unmotivated from the perspective of the informa-
tion processing and underlying mechanism. As we’ve seen, the mechanisms
that precede +T5(2)s are sensitive to environmental features relevant to the
size, shape, and motion relative to shape of the stimulus. They can therefore
support a content that concerns the configuration of visible features. They
do not support the other contents in so far as they are insensitive to wheth-
er the stimulus is nutritious or whether it is a member of a certain species.
No detectors of chemicals nutritious for toads are involved; no detectors of
individuating characters of species are involved. At most, we can say that
the mechanisms that detect the configuration of visible features approxim-
ate as detectors of these.⁸
Natural selection has (as biologists say) satisficed. Other things being
equal, a capacity to tell nutritious things apart from non-nutritious things
would have been good for the toad, as would a capacity to recognize mem-
bers of species that have historically been caught and eaten by them. But the
plain fact is that neither of these capacities evolved. Such capacities might
require fancier cognitive equipment, which is expensive to build, operate,
and maintain, or the needed mutations might have never arisen. Either way,
these are not the capacities that a toad possesses. Instead the toad possesses a
different capacity that merely approximates these.⁹
Of course, it is the capacity possessed and not the capacity approxim-
ated that neuroethologists aim to explain. In part this is a methodological
point. As noted at the end of Section 2.2, neuroethologists need an exact
description of the preferences of the motivated toad before looking for their
neural basis or else they will be looking for the wrong thing. In part, it is
also a logical point. You cannot explain how your car does 30 miles per

⁸ If what’s meant by ‘‘prey’’ is not a member of a species that has historically been
hunted by the toad, but rather something that can be caught and eaten, the suggestion
is much more promising. Not, in my view, because edibility is a Gibsonian affordance,
but because +T5(2)s lead to further information processing that controls catching and
eating, and we do not want to preclude the possibility that +T5(2)s have more content
than their perceptual content (i.e. motor or motivational content).
⁹ One could argue that the toad has a capacity to recognize that food has a certain
probability of being present, but notice that the toad would still not be in error when it
tokened at something with the right configuration of visible features.
186 Karen Neander

gallon if it cannot do 30 miles per gallon and can only do 28 instead. You
could explain how it could do 30 if it were altered in various ways, but
neuroethologists are not primarily interested in explaining how to improve
on the design of creatures like toads.
What of the content toad food that is moving in a direction that parallels its
longest axis? This kind of mixed proposal, intended to be ecumenical, doesn’t
disregard the neuroethology as egregiously as others do, but the compromise
needs motivating. It is a virtue of it that it lets us count a +T5(2) produced
in response to a nutritious ‘‘anti-worm’’ as erroneous. But most of the objec-
tions mentioned above in relation to toad food apply equally well to toad food
with the right configuration of features. The toad does not have a capacity to
recognize whether or not the stimulus is nutritious. It has a capacity to recog-
nize whether the stimulus has the right configuration of features. This latter
capacity approximates the former capacity. But a capacity approximated is
a capacity approximated, not a capacity possessed.

3.3. Responses to Objections

Before closing this chapter, I would like to review a few potential objec-
tions. They cannot all be fully answered here, but it will be as well to
register them and comment on them briefly.
1. Some worry that eating a whatsit would not be rational on the part
of the toad unless the toad thought that whatsits were nutritious (e.g. Price
2001). I do not see this. It is not as if a toad represents a prey-like stimu-
lus as not food or not nutritious. But more importantly, it is controversial
whether the point of making content ascriptions is to rationalize behavior.
That might be the point of making content ascriptions in the context of folk
psychology (Price’s main concern), but rationalizing behavior, as opposed
to explaining it, is not the point of content ascriptions in mainstream cog-
nitive science. It does not assume that all doxastic representational systems
are rational, let alone that all sub-doxastic representational systems are.
Sometimes when this objection is raised it is said that unless the content
ascription rationalizes the behavior, no intentional explanation of the beha-
vior can be given. This depends on what counts as an intentional explan-
ation. However, cognitive scientists can give a representational explanation
of the toad’s behavior, consistent with the content ascription recommended
here. They can explain that the toad is adapted to chase and swallow stimuli
that it represents as having the right configuration of visible features. If that
does not constitute an intentional explanation, so be it.
2. It might be argued that we are forced to select a content like toad food
because an appeal to selection history is the only way to naturalize content
 Content for Cognitive Science 187

and this is the content we get when we appeal to selection history. How-
ever, even if a theory of content must make reference to selection history,
there are different ways to do so, and some theories of content that appeal to
selection history generate contents of the sort favored here, as was noted in
the introduction. Of course, whether they are satisfactory on other grounds
remains to be seen. I am sometimes accused of not being teleological in pro-
posing this argument. However, I see myself as steering us toward a more
promising type of teleological theory.
3. Millikan (2000, app. ) argues that arguments like mine lead us to a
bad end. She thinks that such arguments lead to the conclusion that con-
tent cannot go beyond what can be discriminated. In fact, to my dismay,
Millikan interprets an earlier paper of mine (Neander 1995) as asserting
this, and as even asserting that there can be no distal contents (and no distal
functions)! I did not say this (I certainly did not mean to say this), and
nor am I saying it here. I shall not try to unravel the prior confusion but
we should be clear that the present argument does not entail that content
cannot go beyond what can be discriminated.
Millikan maintains that arguments like mine are ‘‘verificationist’’ (Mil-
likan 2000, app. ). She also levels the claim against Jerry Fodor, for his
having said that, ‘According to informational semantics, if it’s necessary
that a creature can’t distinguish X s from Y s, it follows that a creature can’t
have a concept that applies to X s but not to Y s’ (Fodor 1995: 32.)
Fodor (1990b: 107–9) makes a similar point, in explaining his asym-
metric-dependency theory, which entails that the frog represents (de dicto)
small, black dots rather than flies. Fodor defends this outcome because, he
says, if we think of the content as fly we would have to allow that some
mistakes are nomologically necessary for the frog. He sheds light on what
he means by this when he adds that if we think of the content as fly, ‘There
is no world compatible with the perceptual mechanisms of frogs in which
they can avoid mistaking black dots for flies’ (Fodor 1990b: 107–9).
Fodor refers to the implication as ‘an attenuated sort of verificationism’
(1990b: 108), an admission that Millikan turns into a critique. However,
Fodor’s principle is not really verificationist. For one thing, it applies to
concepts not sentences or sentences in the head. For another, it only applies
to conceptual primitives. Admittedly, most concepts are conceptual primit-
ives according to Fodor, or anyway the Fodor from back then, but we need
to note that we only approach even an attenuated form of verificationism if
we combine Fodor’s principle with this further claim.
Even with respect to conceptual primitives, Fodor’s principle only entails
that their contents cannot go beyond what can be discriminated in a certain
188 Karen Neander

sense. It is unclear how much of a creature’s psychology Fodor wants us to

keep fixed across possible worlds, but a world in which we learn something
that we could learn is obviously consistent with our psychology. This means
that even the contents of conceptual primitives can, according to Fodor’s
principle, go beyond what can presently be discriminated. It allows for a dis-
tinction in referential content between ‘‘water’’ in our mouths and ‘‘water’’
in the mouths of our Twin Earth doppelgängers for instance (see Putnam
1975).
We need to be equally cautious when interpreting my claim that con-
tents have to be supported by relevant information processing. This also
permits contents to go beyond what can be discriminated. Consider the
water–XYZ case again. In 1600 Oscar refers to water, specifically, despite
his inability to tell it apart from XYZ. One popular way to understand this
capacity is that it involves what psychologists call ‘‘essentialist thinking’’.
This in turn seems to involve an intention to defer, in this case, to nature to
delineate the boundaries of the kind term. It seems to involve an intention
to refer to a kind that has a certain Lockean nature, which might be hidden
or unknown, but which is nonetheless expected to explain the superficial
properties by which we recognize instances of the kind as instances of the
kind (e.g. instances of water as instances of water).
According to Frank Keil (1989), human children do not acquire this
capacity until the early years of elementary school, but according to Susan
Gelman and Henry Wellman (1991), there is some evidence of its earliest
manifestations as early as 4 years old. Its development is uniform enough
in children for developmental psychologists to suspect that it is an innate
capacity for our species. I do not know if other primates have this capacity,
but I am very confident that toads do not. It seems a moral certainty that
we have the kind of information processing needed to support essentialist
thinking and that toads do not.
In fact, any theory of mental content that entails that toads have contents
that in us require essentialist thinking is profoundly problematic. It oblit-
erates an important distinction in our capacities versus theirs, and rushes
headlong toward sophisticated contents, heedlessly bestowing them where
they do not belong. There are many degrees of difference between our
brains and those of toads, and a good theory of content needs to observe
them all. Thomson’s gazelles surely do not have essentialist thinking either,
and if so they must lack a water concept or a lion concept that is exactly
like ours. Nonetheless, their perceptual abilities are much more sophistic-
ated than a toad’s. For example, a gazelle can distinguish lion appearances
from the appearances of other predators it encounters. When gazelles see
predators approach, they monitor them, and whether they flee seems to
 Content for Cognitive Science 189

depend on subtle behavioral cues on the predator’s part. How they flee (e.g.
whether they engage in stotting, and how close they let the predator get
before fleeing) depends on the type of predator (Griffin 2001: 71). Further,
they can recognize a lion-like animal from different distances and differ-
ent angles of view, and on the basis of seeing only small parts of it (Griffin
2001: 128). Compare this to the toad’s simple perceptual capacities. The
toad’s detection of prey-like or predator-like stimuli is not mere transduc-
tion, but it is nonetheless based on a fairly simple configuration of visible
features. A good theory of content should respect these gradations in soph-
istication.
4. Another way to respond to the argument in this chapter is to argue
that an innately programmed inference is involved in the case of the toad.
It seems undeniable that the toad is processing information about the con-
figuration of visible features. But it could be argued that the toad’s brain
infers the presence of something nutritious (or the presence of a member
of a prey-species) from this configuration of visible features. According to
most contemporary theories of perception, perceptual processing is often
inferential. It often involves innate assumptions that are ‘‘embodied’’ in the
processing. The relevant inferences are not deliberate or conscious but are
implicit in unconscious processing. So why shouldn’t this be what is hap-
pening in the toad’s case?
The problem with this response is that there is no motivation from the
perspective of neuroethology for thinking that an inference is involved. And
we are looking for pre-theory-of-content reasons for preferring one content
over another. For one thing, if it seems plausible that the toad infers the
presence of nutrients, it should seem equally plausible that it infers the pres-
ence of a member of a prey species, since the configuration of features can
be considered equally good evidence for both. There is nothing in the neur-
oethology that distinguishes between these two competing interpretations.
For another thing, the claim that there is an inference or something like an
inference has empirical implications, if we are realists about representations.
Those who press this reply must allow that some representation in the first
place represents the configuration of features. There must be one represent-
ing the premise (or quasi-premise) and another representing the conclusion
(or quasi-conclusion) of the inference (or quasi-inference). To insist that
one and the same representation represents both premise and conclusion is
an ad hoc move.
Let us take a quick look at a case that cognitive scientists count as
inferential: our perception of parallel lines converging in the distance. It
is thought that we see them as parallel because our perceptual processing
embodies the assumption that lines that converge toward the horizon are
190 Karen Neander

parallel and receding. The idea is that we infer, or sort of pseudo-infer, from
the presence of converging lines in the 2D sketch that parallel and receding
lines are present in the scene, and so our 3D sketch of it represents them
as parallel and receding (Marr 1982). Notice that when cognitive scientists
make this proposal they posit two stages of representation: the 2D sketch
of converging lines and their subsequent 3D interpretation as parallel and
receding. In the case of the toad, there is no evidence of a relevant second
step. Once the toad tokens a +T5(2), it appears to have done all it does by
way of ‘‘prey’’/‘‘predator’’/‘‘other’’ discrimination. From there, it moves on
to processes that govern orienting, approaching, more precise localization
of the prey, and so on.
In any case, the inferential response cannot save certain theories (e.g.
Millikan’s, at least on her construal of it), which do not allow that any
representation of the configuration of visual features occurs. If there is no
representation of the configuration of visual features there can be no infer-
ence from such a representation.
In relation to this, it is also worth noting that a general principle of
information-processing accounts of perception is that visible properties
must be represented before invisible ones are (see e.g. Palmer 1999:
85–92,146–50). What is meant by a ‘‘visible property’’ here is not entirely
clear. However, the idea is that the perceptual system must begin with the
surface features of a perceived object that are presented to the perceptual
system. It must begin with the hide of the cow, not its inner nature, and
with the side of the cow facing the perceiver, not the parts hidden from
view. This principle (often regarded as the downfall of a Gibsonian theory
of perception)¹⁰ makes good sense. And any theory that entails that the
toad’s perceptual system only represents the stimulus as nutritious violates
this principle.
5. Nor does it help to claim that toads need to know about toad food.
Millikan says that mice need to know about hawks, not merely about the
properties by which they recognize hawks (Millikan 2000, app. ). But this
is just plain wrong. The mouse does not need to know about hawks as such.
As long as it escapes, the capacity to recognize hawk-like properties suf-
fices (satisfices). Nor do toads need to know about toad food as such. What
matters is whether they eat the food, not whether they represent it as such.¹¹

¹⁰ I ignore Gibson’s theory of perception in this chapter because it’s not mainstream
cognitive science.
¹¹ This recalls Fodor’s point: ‘All that’s required for frog snaps to be functional is that
they normally succeed in getting the flies into the frogs; and, so long as the little black
dots in the frog’s Normal environment are flies, the snaps do this equally well on either
account of their intentional objects. The mathematics of survival comes out precisely the
 Content for Cognitive Science 191

6. It might also be argued that mainstream cognitive science is just one

approach to understanding animal behavior. I grant this. As I said in the
introduction to this chapter, my conclusions here are conditional on the
assumption that a theory of mental content should meet the needs of main-
stream cognitive science. I am aware that there are radical critiques of
mainstream cognitive science. I am also aware that there might be other sci-
entific purposes, mainstream but in corners of science not examined here,
that pull us in different directions. I leave it to others to argue that these
are strong enough to warrant either pluralism about content or a different
notion of content.
7. Next to last are concerns about distal content. Items with the right
configuration of visual features include worms, and so on, which are distal.
So there is nothing in the choice of content (the configuration of visible
features) that precludes distal content. But does the principle I am advoc-
ating (that the relevant information processing must support the content
we ascribe) drive us toward the proximal? My answer is no, not inexorably.
Elsewhere (in Neander forthcoming) I propose a solution to the problem of
distal content that is sensitive to this principle. (Promises, promises . . . )
8. Finally, it can seem absurd to think that toads have a concept of
something moving in a direction that parallels its longest axis, because this
can seem rather a sophisticated concept. Didn’t we learn the concept of par-
allel lines in school? Is it not more sensible to think that, at most, toads can
possess only basic concepts like food or worm or millipede?
The problem with this objection is that it reverses the relevant ordering
of basic–complex concepts. It would get the order right if what mattered
were the order in which children acquire lexical concepts, but what matters
is what is elementary from the perspective of information processing. Early
visual processing in humans first involves representations of the surface fea-
tures of objects. Recognition of things as food or worms or millipedes is
based on this. The toad, if I am right, does not get that far. Its categories
remain closely tied to the visible features of its environment. The present
proposal does not attribute overly sophisticated capacities to the toad. On
the contrary, compared to its competitors, it attributes less sophisticated
capacities to the toad.
This reminds me of a time when, after I had given a talk at a colloqui-
um, someone began almost yelling at me that his 2-year-old daughter did
not have a concept of an inter-aural disparity. I had been explaining a

same either way’ (Fodor 1990b: 106). As Fodor goes on to say, this is the kind of thing
that makes philosophers feel that content makes no difference, but we have seen that it
makes an explanatory difference.
192 Karen Neander

hypothesis about sound localization. The hypothesis is that, in part, we

hear where a sound is coming from by figuring out which ear received the
pattern of sound first. Since a sound from the right side enters the right
ear a fraction of a second before it enters the left ear, one plausible hypo-
thesis is that the brain can determine the direction of sound (in part) by
determining the difference in the time of its arrival at each ear. It’s not a
straightforward process, because the brain has to figure out which pattern
from one ear matches which pattern from the other. The bird’s trilling has
to be matched with the bird’s trilling, and not with your friend’s talking, of
the fridge’s humming, which might be heard simultaneously. Of course, it’s
not a conscious process. We have no introspective access to the process. We
only have introspective access to its results, when we hear sounds as coming
from a particular direction.
The interjector was confused, despite the impressive volume with which
he explained his point. The hypothesis may or may not require his daugh-
ter to possess a concept of an inter-aural disparity. That depends on what is
meant by ‘‘concept’’. But the proposal does not require her to possess one in
any objectionable sense. People mean many different things by ‘‘concept’’,
but here are two possibilities. According to one, I possess a concept of X just
in case I have introspective access to a semantically structured representa-
tion of Xs. According to the other, I possess a concept of X just in case my
brain employs a representation of Xs. The first is a much more demanding
concept of a concept, and the interjector’s mistake was to think that this
account of sound localization required that his daughter, who could hear
where sounds were coming from, possessed a concept of an inter-aural dis-
parity in the first, demanding sense. Of course, if it requires her to possess
a concept of an inter-aural disparity, it can only require her to possess it
in the second, quite undemanding sense. That is to say, it can only require
that her brain represent inter-aural disparities. And that is not in the least
absurd, for according to plausible psychological theories, she does.
The same sort of thing is true for the toad. There is no suggestion in the
preceding discussion that the toad has any concepts in the first demanding
sense. What is required is merely that its brain has a representation of the
relevant configuration of visible features, and thus a ‘‘concept’’ of them (if
we choose to call it that) only in the second, undemanding sense.¹²

¹² Early drafts were presented to the Department of Philosophy at Duke University

(2004), to the Philosophy of Science Colloquia, University of California at Berkeley
(2002), the Jean Nicod Institute and the École Normale Supérieure, Paris (2002), the
Instituto de Investigaciones Filosóficas, Mexico City (2002), and the Department of
Philosophy, Syracuse University (2001). I benefited a lot from the discussions at these
events and would like to thank the participants.
 Content for Cognitive Science 193

REFERENCES

A, N. (1993), ‘What Do Frogs Really Believe?’, Australasian Journal of Philo-

sophy, 71: 1–12.
B, J. V. Z., and B, L. P. (1962), ‘Experimental Studies of Mimicry,
6. The Reaction of Toads (Bufo terrestris) to Honeybees (Apis mellifera) and their
Dronefly Mimics (Eristalis vinetorum)’, American Naturalist, 96: 297–308.
C, J. (1984), Neuroethology: Nerve Cells and the Natural Behavior of Animals
(Sunderland, Mass.: Sinauer Associates).
C, T. J. (2000), Behavioral Neurobiology (Sunderland, Mass.: Sinauer Associ-
ates).
C, C., and A, M. (1992), ‘Prey-Catching and Predator-Avoidance in Frog
and Toad: Defining the Schemas’, Journal of Theoretical Biology, 157: 271–304.
C, H. B. (1936), ‘The Effectiveness of Protective Adaptation in the Hive Bee,
Illustrated by Experiments on the Feeding Reaction, Habit Formation, and
Memory of the Common Toad (Bufo bufo bufo)’, Proceedings of the Zoological
Society of London (1936), 111–33.
D, D. (1995), Darwin’s Dangerous Idea (New York: Simon & Schuster).
D, F. (1986), ‘Misrepresentation’, in Radu Bogdan (ed.), Belief: Form, Con-
tent and Function (Oxford: Oxford University Press), 17–36.
(1994), ‘If You Can’t Make One You Don’t Know How It Works’, in P. A.
French, T. E. Uehling, and H. K. Wettstein (eds.), Midwest Studies in Philo-
sophy, xix: Philosophical Naturalism (Notre Dame, Ind.: University of Notre
Dame Press), 468–82; repr. with small revisions in D. Chalmers (ed.), Philo-
sophy of Mind (Oxford: Oxford University Press, 2002), 491–9.
E, J. P. (1980), Neuroethology: An Introduction to the Neurethological Funda-
mentals of Behavior, trans. Transemantics Inc. (Berlin: Springer Verlag).
and K, W. (1978), ‘Configural Prey-Selection by Individual Experience in
Toad BufoBufo’, Journal of Physiology, 126: 105–14.
B, H., and S-P, E. (1983), ‘Neuroethological Ana-
lysis of the Innate Releasing Mechanism for Prey-Catching in Toads’, in J. P.
Ewert, R. Capranica, and D. Ingle (eds.), Advances in Vertebrate Neuroethology
(New York: Plenum Press), 413–75.
F, J. (1990a), ‘Psychosemantics, or: Where Do Truth Conditions Come
From?’, in W. Lycan (ed.), Mind and Cognitio: A Reader (Oxford: Blackwell),
312–37.
(1990b), ‘A Theory of Content II’, in Fodor, A Theory of Content and Other
Essays (Cambridge, Mass.: Bradford Books, MIT Press).
(1995), The Elm and the Expert: Mentalese and its Semantics, The Jean Nicod
Lectures (Cambridge, Mass.: MIT Press).
G, S., and W, H. (1991), ‘Insides and Essences: Early Understand-
ings of the Non-Obvious’, Cognition, 38: 213–44; repr. in E. Margolis and S.
Laurence (eds.), Concepts: Core Readings (Cambridge, Mass.: MIT Press, 1999),
613–37.
194 Karen Neander

G, D. R. (2001), Animal Minds: Beyond Cognition to Consciousness (Chicago:

University of Chicago Press).
G, P., C, C., and K, S. K. (1983), ‘Frog Prey Capture Beha-
vior: Between Sensory Maps and Directed Motor Output’, in J. P. Ewert, R.
Capranica, and D. Ingle (eds.), Advances in Vertebrate Neuroethology (New York:
Plenum Press), 331–47.
J, P. (1997), What Minds Can Do: Intentionality in a Non-Intentional World
(Cambridge: Cambridge University Press).
K, F. C. (1989), Concepts, Kinds and Cognitive Development (Cambridge, Mass.,
MIT Press).
K, J. R., and C, C. M. (1996), ‘Visually Elicited Turning Behavior in
Rana Pipiens: Comparative Organization and Neural Control of Escape and Prey
Capture’, Journal of Comparative Physiology, 178/3: 293–305.
L, J. Y., M, H. R., MC, W. S., and P, W. H. (1959),
‘What the Frog’s Eye Tells the Frog’s Brain’, Proceedings of the Institute of Radio
Engineers, 1940–51.
M, D. (1982), Vision (San Francisco: Freeman).
M, R. (1991), ‘Speaking Up for Darwin’, in G. Rey and B. Loewer (eds.),
Meaning and Mind: Fodor and his Critics (Oxford: Blackwell), 151–64.
(2000), On Clear and Confused Ideas: An Essay about Substance Concepts
(Cambridge: Cambridge University Press).
N, K. (1995), ‘Misrepresenting and Malfunctioning’, Philosophical Studies,
79: 109–41.
(forthcoming), Mental Representation: The Natural and the Normative in a
Darwinian World (Cambridge, Mass.: MIT Press).
P, S. (1999), Vision Science: Protons to Phenomenology (Cambridge, Mass.:
MIT Press).
P, D. (1998), ‘Teleosemantics and Indeterminacy’, Australasian Journal of
Philosophy, 76: 1–14.
P, C. (1998), ‘Determinate Functions’, Nous, 32: 54–75.
(2001), Functions in Mind: A Theory of Intentional Content (Oxford: Claren-
don Press).
P, H. (1975), ‘The Meaning of ‘‘Meaning’’ ’, in K. Gunderson (ed.),
Language, Mind and Knowledge (Minneapolis: University of Minnesota Press),
131–93; repr. in Putnam, Philosophical Papers, ii: Mind, Language and Reality
(Cambridge: Cambridge University Press).
S, K. (1990), The Representational Theory of Mind (Oxford: Blackwell).
 9
Representation and Unexploited
Content
Robert Cummins, Jim Blackmon, David Byrd, Alexa Lee,
and Martin Roth

1. INTRODUCTION

In this chapter, we introduce a novel difficulty for teleosemantics, namely,

its inability to account for what we call unexploited content—content a
representation has, but which the system that harbors it is currently unable
to exploit. In Section 2, we give a characterization of teleosemantics. Since
our critique does not depend on any special details that distinguish the vari-
ations in the literature, the characterization is broad, brief, and abstract.
In Section 3, we explain what we mean by unexploited content, and argue
that any theory of content adequate to ground representationalist theories
in cognitive science must allow for it.¹ In Section 4, we show that teleo-
semantic theories of the sort we identify in Section 2 cannot accommodate
unexploited content, and are therefore unacceptable if intended as attempts
to ground representationalist cognitive science. Finally, in Section 5, we
speculate that the existence and importance of unexploited content has
likely been obscured by a failure to distinguish representation from indic-
ation, and by a tendency to think of representation as reference.

¹ There are, of course, initiatives in cognitive science that are not representation-
alist—e.g. the dynamic systems approach advocated by van Gelder and Port (1995)
and others. If non-representationalist approaches ultimately carry the day, then disputes
about how mental representation should be understood in cognitive theory will have
been idle. For the most part, we simply assume in what follows that some form of repres-
entationalism is correct. But, now and again, we phrase matters more methodologically,
as points about what representationalist explanations of cognition need to assume rather
than as points about what cognitive processes actually require.
196 Cummins, Blackmon, Byrd, Lee, and Roth

2. TELEOSEMANTICS

Teleological accounts of representational content identify the extension of

a representation R with the class of things C such that, historically, it was
applications of tokens of R to members of C that led to the selection and
replication of the mechanisms that produced or consumed tokens of R.
Accounts along these general lines are familiar from the writings of Millik-
an, Neander, Papineau, and others (Millikan 1984, 1986; Neander 1991;
Papineau 1984, 1987; recent anthologies by Allen et al. 1998; Buller 1999;
Ariew et al. 2002). For our purposes, the crucial point about all such the-
ories is that a representation R can have the content C for a system only
if it had, when selection took place, the ability to apply R to members of
C. There cannot be selection for an ability that isn’t there. The scenario
underlying teleological accounts of content features a sub-population with a
mechanism (what Cummins 1996a calls an intender) in the R-application
business. It applies R to a variety of things, including, under certain cir-
cumstances, members of C. Those applications—the applications of R to
members of C—prove adaptive enough to cause the mechanism in ques-
tion to spread through the population over time.
It is no part of this story that the reliability of R applications (applica-
tions of R to Cs vs. non-Cs) or their accuracy (excellence of fit between a
token of R and its target C) is ever very good, or improves over time. All
that is required is that it is the applications of R to members of C that leads
to the spread of the mechanism through the population. The trait—the
characteristic pattern of R applications—may go to fixation even though R
is applied somewhat inaccurately to members of C, and only under rather
special or rare circumstances, and frequently applied to other things. We
emphasize this to make it clear that the critique we elaborate in the next
section does not depend on the reliability or accuracy of the selected rep-
resenting or representation-consuming mechanisms. On the other hand,
accurate enough applications of R to Cs cannot simply be random acci-
dents: there must be a mechanism in the R-application business to
select.

3. UNEXPLOITED CONTENT

By unexploited content we mean information or content carried by or present

in a representation that its harboring system is, for one reason or another,
unable to use or exploit. A common-sense example will help to introduce
 Representation and Unexploited Content 197

the basic idea. Imagine someone who learns to use road maps to find a route
from point A to point B. A study of the map might lead to the following
plan: make a left at the third intersection, then another left at the next cross
street, followed by an immediate right. It never occurs to this person to use
the map to extract distance information until, one day, someone suggests that
the map shows a shorter route than the one generated. Prior to this insight,
our imaginary subject uses the map in a way that would be insensitive to vari-
ous geometrical distortions, such as shrinking the north–south axis relative
to the east–west axis. If assignments of representational content are limited
by the abilities its user actually has to exploit the map, we will have to say that
there is no distance information there to be exploited until after the user has
learned to exploit it. And this will evidently make it impossible to explain
how the user could learn to effectively compare routes for length: you can-
not learn to exploit content that isn’t there. Indeed, it is evident that this
story makes no sense at all unless we concede that relative distances are rep-
resented before the user learns to exploit that information. Even if the user
never exploits relative-distance information, we are forced to allow that it is
there to be exploited, since, under the right conditions, the user could have
learned to use maps to compare distances. This would not be possible if the
map did not represent relative distances.
How seriously should we take this sort of example? We think the les-
son is far-reaching and fundamental. To begin with, the idea that a brain
can learn to exploit previously unexploited structure in its representations
is presupposed by all neural network models of learning. Such learning typ-
ically consists in adjusting synaptic weights so as to respond properly to
input activation patterns. This whole process makes no sense unless it is
assumed that input patterns represent proximal stimuli prior to learning,
and that the representational content of input patterns remains the same
throughout learning. Prior to learning, the network cannot properly exploit
input representations: that is precisely what the process of weight adjust-
ment achieves over time.²
Having come this far, we can see that the problem of learning to exploit
‘lower-level’ (‘upstream’) representations must be ubiquitous in the brain, if
we assume that the brain acquires new knowledge and abilities via synaptic
weight adjustment. In perceptual learning, for example, proximal stimuli
must be represented before the appropriate cortical structures learn or
evolve to exploit those representations in target location and recognition.

² We have heard it said that the network creates the content of its input patterns as
learning progresses. But if we say this, we have no reason to say that early responses are
errors. And if early responses are not errors, why change the weights in any particular
direction? Indeed, why change them at all?
198 Cummins, Blackmon, Byrd, Lee, and Roth

Figure 9.1. Depth form texture gradients

As an example, consider the capacity to exploit texture gradients as visual

depth cues. Representations in V1 contain texture gradients but the ability
to exploit these as depth cues (as you do when you view Figure 9.1)
develops later. Similarly, the ability to exploit retinal disparity in binocular
vision as a depth cue develops along with the organization of binocular
columns in the visual cortex. This process can be aborted by exotropy,
but in such cases, binocular fusion without stereoscopic depth vision can
still be achieved as the result of surgical correction and vision training,
demonstrating that the retinal disparity information is still present in early
visual representations, but unexploited for depth.
There is no need to multiply examples. Once our attention is drawn to
the phenomenon, it is clear that there must be many features of represent-
ations, especially structural features, at nearly all levels of perceptual and
cognitive processing, that require learning and/or development for proper
exploitation.

4. TELEOSEMANTICS AND UNEXPLOITED CONTENT

Situating these facts in an evolutionary context immediately reveals a prob-

lem for teleosemantics. It is certainly possible, and probably common, that
the abilities required to exploit various features of representations evolved
well after those features appeared in the representations themselves. As just
remarked, the ability to exploit texture gradients in early visual representa-
tion as depth cues might well have evolved well after well-defined gradients
were available in those early representations. Now here is the point: the
presence of texture gradients in early visual representations could not have
been adaptive prior to the evolution of the processes that exploit them.
Teleosemantics, however, implies that texture gradients did not represent
depth until after it became adaptive for visual representations to include
 Representation and Unexploited Content 199

them. In general, content only becomes adaptive, hence a candidate for the
kind of content-fixing selection contemplated in teleosemantics, when and
if the ability to exploit it is acquired. Evidently, there can be no selection
for an ability to exploit content that isn’t there. The ‘opportunity’ to evolve
the ability to exploit texture gradients in visual representations as depth cues
simply cannot arise unless and until depth-representing texture gradients
become available to exploit.³
Reflection on this last point suggests that the same difficulty arises
whether the ability to exploit some feature of a representation is learned
or evolved. For concreteness, assume that the ability to exploit texture
gradients as depth cues is learned. While the current state of neuroscience
provides no definitive account of such learning, it is perfectly intelligible to
suppose it involves the systematic adjustment of synaptic weights in some
substructure of the visual cortex. Evidently, if the ability to learn to exploit
texture gradients itself evolved after texture gradients became available in
early visual representations, we have a situation exactly like the one just
rehearsed: teleosemantics assigns no content to unexploited features of
representations, and this undermines the obvious explanation of how the
ability to learn to exploit such features might later become adaptive.
To sum up: once our attention is directed to the phenomenon of un-
exploited content, it is natural to ask how the ability to exploit previously
unexploited content might be acquired. Learning in the individual, and
evolution in the species, are the obvious answers. Equally obvious, how-
ever, is that teleosemantics cannot allow for evolving the ability to exploit
previously unexploited content: that requires content to pre-date selection,
and teleosemantics requires selection to pre-date content.

5. REPRESENTATION AND INDICATION

It seems likely that the very possibility of unexploited content has been
overlooked in philosophical theories of content because of a failure to dis-
tinguish representation from indication. In this section, we digress a bit to
explain how we understand this distinction, and conclude by suggesting
how exclusive attention to indication tends to make the phenomenon of
unexploited content difficult to discern.⁴

³ The underlying general point here, that selection for a given capacity requires that
the capacity already exist in some part of the population, is not new. See e.g. Macdonald
(1989).
⁴ This section draws heavily from Cummins and Poirier (2004).
200 Cummins, Blackmon, Byrd, Lee, and Roth

5.1. Terminology
Some authors (e.g. Schiffer 1987) use ‘‘mental representation’’ to mean any
mental state or process that has a semantic content. On this usage, a belief
that the Normans invaded England in 1066 counts as a mental representa-
tion, as does the desire to be rich. This is not how we use the term. As we
use the term, a mental representation is an element in a scheme of semantic-
ally individuated types whose tokens are manipulated—structurally trans-
formed—by (perhaps computational) mental processes. Such a scheme
might be language-like, as the Language of Thought hypothesis asserts
(Fodor 1975), or it might consist of (activation) vectors in a multidimen-
sional vector space as connectionists suppose (e.g. Churchland 1995). Or
it might be something quite different: a system of holograms, or images,
for example.⁵ An indicator, on the other hand, simply produces structurally
arbitrary outputs that signal the presence or magnitude of some property in
its ‘receptive field’.

5.2. Indication
We begin with some influential examples.
• Thermostats typically contain a bimetallic element whose shape indicates
the ambient temperature.
• Edge detector cells were discovered by David Hubel and Torsten Wiesel
(1962). They write: ‘The most effective stimulus configurations, dictated
by the spatial arrangements of excitatory and inhibitory regions, were
long narrow rectangles of light (slits), straight-line borders between areas
of different brightness (edges), and dark rectangular bars against a light
background.’
• ‘Idiot lights’ in your car come on when, for example, the fuel level is low,
or the oil pressure is low, or the engine coolant is too hot.
‘‘Indication’’ is just a semantic-sounding word for detection. Since we
need a way to mark the distinction between the mechanism that does the
detection, and the state or process that is the signal that the target has been
detected, we will say that the cells studied by Hubel and Wiesel are indic-
ators, and that the pattern of electrical spikes they emit when they fire are

⁵ It is possible that the brain employs several such schemes. See Cummins (1996b)
and Cummins et al. (2001) for further discussion of this possibility.
 Representation and Unexploited Content 201

indicator signals. Similarly, the bimetallic element found in most thermo-

stats is an indicator, and its shape is the signal.

5.3. Indication vs. Representation

Indication is generally regarded as a species of representation. Indeed, caus-
al and informational theories of representational content assert that rep-
resentation is, or is inherited from, indicator content.⁶ We think the two
should be kept distinct.
Indication is transitive, representation is not. If S3 indicates S2, and S2
indicates S1, then S3 indicates S1. Imagine a photo-sensitive cell pointed at
an ‘idiot light’ in your car, and attached to a relay activating an audio device
that plays a recording: ‘The oil pressure is low.’ If the light indicates low oil
pressure, so does the recording. Representation, on the other hand, is not
transitive. A representation of the pixel structure of a digitized picture of the
Statue of Liberty is not a representation of the statue’s visual appearance,
though the later may be recovered from the former.⁷ To anticipate some
terminology we will use later, a representation of the pixel structure is an
encoding of the statue’s visual appearance.⁸
Indicator signals are arbitrary; representations are not. This is implied by
the transitivity of indication. Given transitivity, anything can be made to
indicate anything else (if it can be detected at all), given enough ingenuity
and resources. Because indicator signals are arbitrary, disciplined structural
transformations of them cannot systematically alter their meanings. Such
transformations, however, are precisely what make representations useful.
Consider, for example, a software package that takes a digitized image of
a face as input and ‘ages’ it, i.e. returns an image of that face as it is likely
to look after some specified lapse of time. Nothing like this could possibly
work on an input that was required only to indicate a certain face—a name,

⁶ The theory is generally credited to Denis Stampe (1977). Its most prominent
advocates are Fodor (1987) and Dretske (1981).
⁷ Representation, on the view advocated by Cummins (1996a), is grounded in
isomorphism. Since isomorphism is plainly transitive, it might seem that representation
must be transitive too. In a sense, this is right: the things that stand in the isomorphism
relation are structures—sets of ‘objects’ and relations on them. If S1 is isomorphic to S2,
and S2 is isomorphic to S3, then S1 is isomorphic to S3. An actual physical representation,
however, is not an abstract object; it has a structure—actually, several—but it isn’t itself
a structure. The connected graph structure of a paper road map is isomorphic to the
street and intersection structure of a town, but not to the town’s topology. The town’s
topology is isomorphic to the topology of a citrus grove. But no structure of the road
map need be isomorphic to any structure of the grove.
⁸ It is what Haugeland would call a recording of the picture. See Haugeland (1990).
202 Cummins, Blackmon, Byrd, Lee, and Roth

say—because there is no correlation between the physical characteristics

something must have to be a signal that indicates the appearance of a face at
age 18 and the physical characteristics of that face at age 18. It follows from
the nature of indication that the structural properties of an indicator signal
have no significance. Indicators ‘say’ that their targets are there, but do not
‘say’ anything about what they are like. Representations, on the other hand,
mirror the structure of their targets (when they are accurate), and thus their
consumers can cognitively process the structure of the target by manipulat-
ing the structure of its representation. But representations, unlike indicator
signals, are typically silent concerning whether their targets are ‘present’:
they are not, except incidentally and coincidentally, detector signals.
Indicators are source-dependent in a way that representations are not. The
cells studied by Hubel and Wiesel all generate the same signal when they
detect a target. You cannot tell, by looking at the signal itself (the spike
train), what has been detected. You have to know which cells generated
the signal. This follows from the arbitrariness of indicator signals, and is
therefore a general feature of indication: the meaning is all in who shouts,
not in what is shouted.⁹
In sum, then, indication is transitive, while representation is not. It fol-
lows from the transitivity of indication that indicator signals are arbitrary
and source-dependent in a way in which representations are not, and this
disqualifies indicator signals as vehicles for structure-dependent cognitive
processing. Representation is intransitive, non-arbitrary, and portable (not
source-dependent), and therefore suitable for structural processing. Indic-
ator signals ‘say’ their targets are present, but ‘say’ nothing about them;
representations provide structural information about their targets, but do
not indicate their presence. Indicator signals say ‘My target is here’, while
representations say ‘My target, wherever it is, is structured like so’.

5.4. Discussion
If indication is your paradigm of mental content, as it is bound to be if
you hold some form of causal theory, you are going to focus on what
fixes the content of an indicator signal.¹⁰ Whatever fixes the content of an
indicator signal, it is not its structural properties. In this context, therefore,
motivation is lacking for thinking about which aspects of a representation’s

⁹ We do not mean to imply here that the shape of a spike train is never significant.
The point is rather that two indicators can have the same spike train, yet indicate different
things.
¹⁰ See Cummins (1997) for more on the marriage between causal theories, indication,
and the Language of Thought.
 Representation and Unexploited Content 203

structure can usefully be processed, and whether the ability to do that pro-
cessing is learned or evolved or a combination of both. Maps rub your nose
in the possibility of unexploited content; idiot lights do not.
There can, however, be unexploited indicator signals. Think of the color-
coded idiot lights at intersections: you have to learn that red means stop,
green means go. Before learning, this is also unexploited content (though
not what we have been calling representational content), and, unsurpris-
ingly, it makes trouble for teleosemantics. Teleosemantics implies that an
indicator signal has no content until there has been selection for the indic-
ator that generates it. But the ability to exploit, or to learn to exploit, an
indicator signal can only evolve if the indicator is already there signaling its
target.
Magnetosomes are magnetically polarized structures (typically ferrite
surrounded by a membrane) in single-cell ocean-dwelling anaerobic
bacteria. The orientation of these structures correlates with the direction
of the earth’s magnetic field. By following the magnetic orientation in a
particular direction, organisms far from the equator can avoid aerobic water
near the surface. For this to work, magnetosomes must be chained and
attached at both ends of the cell to form a reasonably straight line in the
direction of locomotion (see Figure 9.2). This is because the orientation
of the organism is simply a consequence of the orientation of the chain
of polarized molecules. The whole body of the bacterium is a floating
compass needle. The organism swims, and will move in whatever direction
it happens to point.
Chaining, of course, is simply a physical consequence of having a lot of
little magnets suspended in proximity. They will stick together north to

Figure 9.2. Magnetotactic bacterium from the Chiemsee, Bavaria, Germany (Bio-
magnetism Group, University of Munich). Dark blobs are sulfur granules
204 Cummins, Blackmon, Byrd, Lee, and Roth

south. What is not so obvious is why the north pole of the string winds up
attached at the ‘front’—i.e. direction of locomotion—end of the organ-
ism. However this happens, it is certainly possible, indeed probable, that
the attachment process evolved after magnetosomes themselves appeared
within the cell body of anaerobic bacteria. Selectionist theories imply that
magnetosome chains did not indicate the direction of anaerobic water until
after it became adaptive to do so, i.e. only after the evolution of attach-
ment. But surely it is in part because they did indicate the direction of
anaerobic water that the attachment process was adaptive enough to be
selected for.

6. CONCLUSION

A very natural response to the foregoing is to say that unexploited content

isn’t really content. After all, there is a lot of unexploited information in the
environment, information that cognitive systems must acquire the abilities
to exploit. We do not call that information content.
We are sympathetic with the comparison between learning or evolving
an ability to exploit information in a representation or indicator signal and
learning or evolving an ability to exploit information in the environment.
We think these are, in fact, deeply similar. The importance of this simil-
arity is obscured or lost in theories that essentially take representation to
be reference. Theories of content that take representation to be reference
perforce focus on devising the conditions that (allegedly) fix the references
of semantically primitive terms, relying on the standard truth-conditional
combinatorics to fix the references and truth-conditions of complex expres-
sions. Access to something that refers to horses—a primitive term in
Mentalese—however, tells you nothing about horses. Actual information
about horses, therefore, is to be found only in the (or a) set of Mentalese
sentences that contain a |horse| (a Mentalese term referring to horses) and
appear in the Belief Box. The only sense in which such an account allows
for unexploited content, therefore, is the sense in which a cognitive agent
might not exploit all of its beliefs about horses on a particular occasion.
While this is undoubtedly a case of unexploited information, it is not a case
of the sort we have been discussing. Returning to our analogy, inability to
extract relative-distance information from a road map is quite distinct from
failing to read or utilize some of the sentences in a book. In the later case,
the content of the unread sentences is unproblematically extractable from
those sentences; they just are not noticed for one reason or another. The
problem is not that one doesn’t know how to read them. In the case of the
map, a new skill is required to exploit the needed information. Unexploited
 Representation and Unexploited Content 205

information of the sort allowed for in Language of Thought (LOT) theories

evidently poses no problem for teleosemantics comparable to the one we
have been urging, since the mechanisms responsible for applying the prim-
itive constituents and processing the relevant syntactical machinery may be
selected for independently of their occurrence in any particular belief.
Cognitive systems need information. LOT accounts attempt to provide
for this by giving indication a twofold role. First, indicator signals alert
the organism to the instantiation of their target properties in their recept-
ive fields.¹¹ Second, primitive terms of LOT inherit their references from
the properties they are used to indicate in the context of detection. Cog-
nition is understood as the application of theories expressed as organized
sets of sentences in Mentalese, hence as a species of truth-conditional infer-
ence, implemented as computation over symbolic structures with Tarskian
logical forms.
Perhaps something like this story makes sense for the ‘higher’ cognition
of adult humans, especially if, like Plato, one is inclined to suppose that
cognition is thinking and that thinking is essentially talking to oneself. But
this picture is of little use for understanding phenomena like the capacity to
exploit texture gradients in early visual representations as depth cues. When
we turn to phenomena such as these, the truth-conditional semantics of
propositional attitude contents is of dubious significance to cognitive sci-
ence. Much more important, we believe, is the conception of representation
and indication briefly introduced above. Representations, thus conceived,
are of use to cognitive systems as an information source in addition to the
environment (i) because they can be stored, and (ii) because they can be
structurally transformed in ways that the environment typically cannot be.
Sophisticated indicators are of use because they can signal the presence of
environmental features—e.g. the presence of a predator or a mate—that
are extremely abstract from the point of view of proximal stimulation. Rep-
resentation and indication thus conceived are what make reference and
the propositional attitudes possible. A theory that begins with the truth-
conditional semantics of propositional attitude contents thus skips over
most of the action and begins with a phenomenon that is just the sort of
cognitive achievement that mainstream cognitive science seeks to explain in
terms of representation.
We do not wish to quibble over whether the phenomenon we have called
unexploited content is really content. We do contend, however, that what
we are calling content is what ultimately does the work in representationalist

¹¹ See Cummins and Poirier (2004) for a discussion of how indicators might become
‘source-free’ and function as terms.
206 Cummins, Blackmon, Byrd, Lee, and Roth

cognitive science.¹² No doubt we need to mark the distinction between

exploited and unexploited content. We think ‘‘exploited content’’ and
‘‘unexploited content’’ do the job nicely. Refusing to use the word ‘‘con-
tent’’ for as yet unexploited features of structured representations strongly
suggests, wrongly, that those features are somehow different from those that
are exploited. There is no intrinsic difference between texture gradients that
are exploited and texture gradients that are not. To suppose otherwise would
be like supposing that road maps cease to represent relative distances in the
hands of those who cannot extract that information from them.¹³
In this chapter, we have urged what we think is a novel objection to
teleosemantic theories, namely that they cannot accommodate unexploited
content or information. Surely, a necessary condition for the plausibility of
a theory of mental representation that hopes to ground representationalist
cognitive science is that it accommodate unexploited content or informa-
tion. For it must be possible for a system to be able to learn or evolve the
capacity to exploit the information carried by a representation or indicator
signal, and this implies that the information is there prior to acquisition of
the capacity to exploit it.

REFERENCES

A, C., B, M., and L, G. (eds.) (1998), Nature’s Purposes (Cam-
bridge, Mass.: MIT Press).
A, A., C, R., and P, M. (eds.) (2002), Functions: New Essays in
the Philosophy of Psychology and Biology (Oxford: Oxford University Press).
B, D. (ed.) (1999), Function, Selection, and Design (New York: SUNY Press).
C, P. (1995), The Engine of Reason, the Seat of the Soul (Cambridge,
Mass.: MIT Press).

¹² An anonymous reviewer complained that we have not cited actual examples of

cognitive scientists appealing to unexploited content. We are saying that they presuppose
it whenever they assume that a representational capacity or ability is learned or evolved.
Presumably, they were all learned or evolved.
¹³ There is a temptation to think that an unexploited feature of a representation
doesn’t represent anything to (or for) the system that harbors it. This is surely right.
But to assimilate representation to representation to/for will, like teleosemantics, make it
impossible to understand how, for example, the ability to exploit texture gradients as
depth cures could be learned or evolved. For more on the representation/representation-to
distinction, see Cummins (1996a).
Notice, by the way, that what is important about texture gradients is not just that they
somehow covary with depth. It is their suitability for structural processing that makes
them useful. When covariation is all that matters, an arbitrary indicator signal is all that
is required.
 Representation and Unexploited Content 207

C, R. (1996a), Representations, Targets, and Attitudes (Cambridge, Mass.:

MIT Press).
(1996b), ‘Systematicity’, Journal of Philosophy, 93: 591–614.
(1997), ‘The LOT of the Causal Theory of Reference’, Journal of Philosophy,
94: 535–42.
and P, P. (2004), ‘Representation and Indication’, in H. Clapin,
P. Staines, and P. Slezak (eds.), Representation in Mind (New York: Elsevier).
B, J., B, D., P, P., R, M., and S, G. (2001),
‘Systematicity and the Cognition of Structured Domains’, Journal of Philosophy,
98: 167–85.
D, F. (1981), Knowledge and the Flow of Information (Cambridge, Mass.:
MIT Press).
F, J. (1975), The Language of Thought (New York: Thomas Y. Crowell).
(1987), Psychosemantics: The Problem of Meaning in the Philosophy of Mind
(Cambridge, Mass.: MIT Press).
H, J. (1990), ‘Representational Genera’, in W. Ramsey, S. Stich, and
D. Rumelhart (eds.), Philosophy and Connectionist Theory (Hillsdale, NJ: Law-
rence Erlbaum).
H, D., and W, T. (1962), ‘Receptive Fields, Binocular Interaction and
Functional Architecture in the Cat’s Visual Cortex’, Journal of Physiology of
London, 160: 106–54.
M, G. (1989), ‘Biology and Representation’, Mind and Language, 4:
186–200.
M, R. (1984), Language, Thought, and Other Biological Categories (Cam-
bridge, Mass.: MIT Press).
(1986), ‘Thoughts without Laws: Cognitive Science with Content’, Philo-
sophical Review, 95: 47–80.
N, K. (1991), ‘The Teleological Notion of Function’, Australasian Journal
of Philosophy, 69: 454–68.
P, D. (1984), ‘Representation and Explanation’, Philosophy of Science, 51:
550–72.
(1987), Reality and Representation (Oxford: Blackwell).
S, S. (1987), Remnants of Meaning (Cambridge, Mass.: MIT Press).
S, D. (1977), ‘Towards a Causal Theory of Linguistic Representation’, in
P. A. French, T. E. Uehling, and H. K. Wettstein (eds.), Midwest Studies in
Philosophy, ii: Studies in the Philosophy of Language (Minneapolis: University of
Minnesota Press).
 G, T. J., and P, R. (1995), ‘It’s About Time: An Overview of the
Dynamical Approach to Cognition’, in R. Port and T. van Gelder (eds.), Mind
as Motion: Explorations in the Dynamics of Cognition (Cambridge, Mass.: MIT
Press).
 10
Fearing Fluffy: The Content
of an Emotional Appraisal
Carolyn Price

1. INTRODUCTION

Harry was visiting his friend Mel. He had wandered into the kitchen to
make a cup of tea, when he heard a sound behind him. He swung round
to see Mel’s cat Fluffy hissing at him from the floor. Seeing Fluffy’s aggres-
sive stance, Harry panicked: he kicked out at the cat and bolted from the
kitchen.
Later, Harry described to Mel what happened: ‘I evaluated the situation’,
he said, ‘and judged that Fluffy presented a danger.’ Harry, it seems, was
not being entirely honest: there is a difference between the dispassionate
evaluation that he described to Mel and the emotional response that he
actually produced. But what could the difference be? The problem is par-
ticularly puzzling if we accept that emotions are not simply bodily feelings
or urges, but involve appraisals of some kind. How can we distinguish emo-
tional appraisals from dispassionate evaluations?
There are a number of ways in which the distinctive character of an
emotional appraisal might be explained. One strategy is to suggest that
emotional appraisals are psychological states of a distinctive type. They are
not judgements, thoughts, desires, or perceptual states; nor are they some
combination of these. Rather, they are produced by a separate psychological
mechanism, and have a special role to play in our psychology (Griffiths
1990). Again, it might be suggested that emotional appraisals have a special
kind of content.¹ For example, it is sometimes suggested that the content of
emotional appraisals is related in a distinctive way to the subject’s interests
or concerns (Averill 1980: 310; Nussbaum 2001: 52 n.).

¹ I use the term ‘content’ to refer to the truth (or correctness) conditions or the
satisfaction conditions of intentional states.
 The Content of an Emotional Appraisal 209

These strategies might be seen as alternatives; but they are not mutually
exclusive. It may be that a complete explanation of the distinctive character
of emotional appraisals will refer to their origin, their function, and their
content. Indeed, it is reasonable to expect these considerations to be linked.
In what follows, I would like to explore these links. In particular, I would
like to explore the idea that the content of emotional appraisals is depend-
ent on their function. I shall concentrate on a single case study: Harry’s
appraisal of the danger presented by Fluffy. I shall refer to this appraisal
as AP .
My discussion will assume a certain theoretical background: that is, a
teleosemantic theory of intentional content of the kind suggested by Ruth
Millikan (1984, 1989). A theory of this kind begins from the claim that the
content of an intentional state is determined by its biological function; or,
more precisely, by the biological function of the mechanism that produced
it, together with the way in which that mechanism normally works. Mil-
likan’s version of the theory is marked by two distinctive features. First,
on her approach, the notions of a biological function and of normality
are understood historically.² Secondly, on her account, we cannot determ-
ine the content of an intentional state by considering only the conditions
under which it would normally be produced: we must also consider the
way in which a state of that type would normally benefit the subject (Mil-
likan 1984: 100; 1989: 286; Price 2001: 82). I shall follow Karen Neander
in referring to a theory of this type as a High Church teleosemantic theory
(HCT) (Neander 1995).
Elsewhere, I have attempted to apply a version of HCT to the content of
relatively simple intentional occurrences, such as sensory signals; and I have
attempted to extend that account to increasingly sophisticated intentional
devices, including perceptual states with singular content, spatial maps,
judgements, and desires (Price 2001). In this chapter, I shall rely on some of
the claims and distinctions that I introduced in my earlier discussion. I shall
introduce these bits of baggage, as briefly as I can, as I go along.
In order to present a teleosemantic account of the content of emotional
appraisals, it will be necessary to say something about the function that
these states play in our psychology. This is, of course, a highly controver-
sial issue; it is also an empirical issue, on which philosophers are entitled to
do no more than speculate. In the next two sections, I shall offer a spec-
ulative account of the function of certain emotional appraisals, drawing

² Roughly, an item can be described as having the function to perform a certain

task if its presence is explained, in part, by the fact that earlier items of the same type
successfully performed that task (Millikan 1984: 25–9; 1989); an item will be carrying
out its function in a normal way if its presence is explained, in part, by the fact that
earlier items performed that task in the same way (Millikan 1984: 33).
210 Carolyn Price

on suggestions made by a number of philosophers and psychologists. The

claims that I shall go on to make about the content of AP must be regarded
as conditional on the truth of this account.

2. WHAT IS AN EMOTIONAL APPRAISAL?

I shall begin with some brief remarks aimed at clarifying what I mean by
an emotional appraisal. I take it that the occurrence of an emotion is a
complex event, involving a structured pattern of physiological and psycho-
logical changes, triggered by an intentional state of some kind.³ I use the
term ‘emotional appraisal’ to refer to the intentional state that triggers these
changes.⁴
In this chapter, I shall assume that the changes involved in a particular
occurrence of emotion are triggered by a single appraisal. This assumption
may well turn out to be false. It has been suggested that an occurrence of
emotion involves at least two processes: a speedy, automated process and
a slower process that involves higher-level cognitive capacities (Oatley and
Johnson-Laird 1987; LeDoux 1998; Toates 2002). If so, it is possible that
these processes are initiated by two or more distinct appraisals, each per-
forming a different set of functions. If this turns out to be correct, it will be
necessary to adjust my account in order to accommodate this point.

3. GENERAL- AND SPECIAL-PURPOSE SYSTEMS

In this section, I shall introduce a distinction that may help us to under-

stand the function of an emotional appraisal.
Human beings appear to possess a psychological system that is involved
in making reasoned decisions about what to do in any given situation. This
reasoning system is capable of drawing on a range of intentional states,
including factual judgements, evaluations, and desires. Millikan suggests
that there is an important distinction between this system and other inten-
tional systems. Our reasoning system is a general-purpose system: it can trig-
ger behaviour that serves almost any interest that we possess or that we

³ For two different views of what might be included within an occurrence of emotion,
see Ekman (1980: 80–95); Goldie (2000: 12–16).
⁴ For present purposes, I shall sidestep questions about how these states are produced,
leaving it open whether they are generated by a separate psychological mechanism or
whether they are triggered by beliefs or desires. If HCT is correct, a complete account of
the content of emotional appraisals will need to take account of the way in which they
are produced; but I do not have space to explore this issue properly here.
 The Content of an Emotional Appraisal 211

may acquire during the course of our lives (Millikan 1986: 72; Price 2001:
191–211).
General-purpose systems benefit from a capacity to deploy information
and practical skills in a highly flexible way. There is no telling in advance
how any of the intentional states produced by the system will be used in
pursuing the subject’s interests. On the other hand, systems of this kind
face some significant challenges. First, because a general-purpose system
serves a wide range of different interests, it needs to have some means
of determining which interest to pursue in any given situation. Secondly,
general-purpose systems have access to a broad range of information, only a
small subset of which is relevant in any given situation. The system must
have some way to ensure that the right information is brought to bear
on the problem. A similar problem arises with respect to the behavioural
responses that the system is able to deploy: the system must ensure that the
responses that it considers are likely to be of some use. These problems are
particularly pressing in situations that need to be dealt with quickly.
General-purpose systems contrast with special-purpose systems. In par-
ticular, they contrast with intentional systems that are interest-dependent:
these are systems that serve a specific interest or set of interests—for ex-
ample, predator-avoidance or nest-building. Interest-dependent systems
range from simple recognition–response mechanisms to systems with relat-
ively sophisticated inferential capacities. The information and behavioural
responses available to an interest-dependent system are used to promote
only the interest or set of interests that the system serves. As a result, organ-
isms that possess only interest-dependent intentional systems will be able to
make only limited use of the information and skills that they acquire.
Interest-dependent systems avoid some of the problems that confront
general-purpose intentional systems. In particular, a system that serves only
a single interest will not need to decide which interest to serve in any
given situation. Moreover, such a system will not face the same degree
of difficulty in focusing on relevant information and effective behavioural
responses: a system of this kind will normally have access only to informa-
tion that is likely to help it to execute the task that it functions to execute;
moreover, the behavioural responses that it is able to deploy will normally
be limited to responses that have, in the past, proved effective in enabling it
to execute this task.
The distinction between interest-dependent and general-purpose
systems bears some similarities to Fodor’s distinction between modular
and non-modular systems. Fodor characterizes modules as systems that
are both domain-specific and informationally encapsulated: that is, they
have a specific informational or executive task to perform and are able
to draw on information from only a limited set of sources (Fodor 1983:
212 Carolyn Price

47–52, 64–86). Like modular systems, interest-dependent systems are

domain-specific. However, they need not be informationally encapsulated:
as we have seen, they will normally have access only to information that
is relevant to the interests that they serve; but this does not, by itself,
rule out the possibility that such a system might call on a wide array of
informational sources, including intentional states generated by a general-
purpose system. The distinction between general-purpose and interest-
dependent systems focuses on the way in which the states that they generate
are put to use, rather than the way in which they are produced.

4. THE FUNCTION OF AN EMOTIONAL APPRAISAL

With this distinction at the ready, we can now turn our attention to the
first question that we need to address: what is the biological function of an
emotional appraisal?
It is not axiomatic that emotional appraisals have a biological function
at all. For this to be case, one of two possibilities must be realised. One
possibility is that we have inherited a set of emotional capacities that once
helped our ancestors to survive and to reproduce, and which thereby help to
explain our presence today. If so, our emotional capacities will function to
help us in just the way in which they helped our ancestors.
Secondly, it is possible that our emotional capacities are not inherited,
but develop during the course of our lives, in a way that depends on our
experiences as individuals and as members of a certain social group. It
might be thought that this scenario is incompatible with the claim that our
emotional capacities have a biological function. But the two claims can be
made compatible if we assume that the process of learning is controlled by
mechanisms that themselves have a biological function. If such mechanisms
exist, their function will be to generate emotional capacities suited to our
physical and social environment. On this scenario, our emotional capacities
could be ascribed functions deriving from the function of these mechan-
isms, together with the way in which they normally work (Millikan 1984:
46–7; Price 2001: 124–9). In what follows, I shall assume that one of these
possibilities is realised. I shall not try to adjudicate between them.⁵
I shall begin from the suggestion that the function of emotions is
to enable us to deal with what Paul Ekman calls fundamental life tasks
(Ekman 1992: 171). These include hazards, such as an encounter with a
large predator; and opportunities, such as an encounter with a potential

⁵ For some different views, see Ekman (1980, 1992); Griffiths (1997); Averill (1980);
Goldie (2000: 84–122).
 The Content of an Emotional Appraisal 213

mate. These situations are crucial for the subject, in the sense that an
inappropriate response can be highly damaging. In addition, they are often
emergencies, to which it is necessary to respond very quickly. According to
the life task hypothesis, the function of emotional appraisals is to enable us
to deal effectively with situations of these kinds.⁶
How do emotional appraisals help us to deal with these situations?
Emotional appraisals are produced quickly, enabling the subject to identify
the situation without delay. They trigger physiological changes, preparing
the subject for action; and they prompt expressive behaviour, signalling
to others how the subject is likely to react. In addition, they generate
behaviour that is designed to resolve the situation: for example, fleeing from
the threat; avenging the insult; celebrating the goal. Unlike some of the
expressive behaviours triggered by emotional appraisals, these behaviours
are not plausibly regarded as stereotypical responses: they are generated by
practical inference. This implies that emotional appraisals are capable of
influencing practical decision-making in some way.
How do emotional appraisals influence decision-making? First,
emotional appraisals are plausibly regarded as sources of motivation.
Moreover, emotional motivations are urgent. In other words, they do not
compete on an equal footing with the subject’s other goals. Gripped by
fear, Harry would not normally balance his desire to avoid being injured by
Fluffy against his desire to make himself a cup of tea: his panicky appraisal
prompts him to treat the goal of protecting himself from Fluffy as his only
current concern.
Secondly, emotions seem to influence the way in which we think. We
saw earlier that a subject who is capable of general-purpose reasoning faces
a fundamental difficulty: they must ensure that, in deciding how to act in a
particular situation, they focus primarily on considerations that are relevant
to the situation. Ronald de Sousa suggests that emotions provide a solution
to this problem: emotions function to frame our reasoning, by focusing our
attention on information that is relevant to the problem at hand (de Sousa
1987: 190–6; see also Evans 2002). If this proposal is correct, then one
function of AP is to fix Harry’s attention on the threat posed by Fluffy and
on any aspect of the situation that might help him to escape or to ward off
the threat.
De Sousa’s proposal concerns the information that the subject considers
before acting. We saw earlier that a similar difficulty arises with respect to

⁶ See also Griffiths (1990); Tooby and Cosmides (1990); Lazarus (1991); Johnson-
Laird and Oatley (1992); LeDoux (1998). Theorists who have argued for the life task
hypothesis have generally supposed that our emotional capacities are inherited; however,
it would be possible to combine the life task hypothesis with the view that our emotional
capacities are learned.
214 Carolyn Price

the possible actions that the subject will consider. Oatley and Johnson-
Laird suggest that a further function of an emotional appraisal is to focus
the subject’s attention on a specific set of responses to the situation
(Oatley and Johnson-Laird 1987: 37). This seems plausible. If we are told
that Harry is overcome by fear, we would expect him not only to try to
avoid being injured by Fluffy, but to do so in certain predictable ways—for
example, by running away from her, by trying to fend her off, or by cower-
ing in a corner. We would not expect him to react by talking calmly to
her, even if this is known to be an effective way to deal with angry cats. To
respond by talking gently to Fluffy would require Harry to master his fear,
not to act in accordance with it. This lends support to the suggestion that
AP works to narrow down the types of action that Harry is able to consider.
De Sousa sums up his proposal by suggesting that the function of emo-
tions is to cause the mechanisms that are responsible for practical reas-
oning to operate as if they were informationally encapsulated modules. I
would like to suggest instead that one function of an emotional apprais-
al is to cause our general-purpose reasoning system to act as if it were a
special-purpose, interest-dependent system. As we saw earlier, such a system
will focus exclusively on a single task; it will normally have access only to
information that is relevant to that task; and it will be able to generate a
limited set of actions, each of which has already proved effective in dealing
with that task.
All this suggests that an emotional appraisal has a very complex function.
We might describe it roughly as follows. In a situation in which the subject
is confronted by a certain type of crucial situation, an emotional appraisal
functions:
1. to prompt the subject immediately to find a way to resolve the situation,
without regard to other considerations;
2. to focus the subject’s attention on a narrow range of possible actions;
3. to focus the subject’s attention on information that will help him or her
to choose one of these possible actions and to perform it in an effective
way;
4. to trigger physiological changes that prepare the subject to carry out one
of those solutions;
5. to trigger expressive behaviour that signals the subject’s situation and
likely actions to other organisms.
How plausible is this account? Certainly it is not complete: it ignores,
for example, the longer-term effects of our emotional experiences on mood,
motivation, and memory. Moreover, this account does not fit all types of
emotional appraisal equally well. In particular, clause (2) does not apply to
 The Content of an Emotional Appraisal 215

all kinds of emotional appraisal. This becomes clear if we contrast Harry’s

panicky response to Fluffy with a case of anxiety. Suppose that Harry is
anxious because he has been told that he is suffering from a dangerous ill-
ness. Knowing that Harry was anxious about the news would not tell us
anything about how he might seek to deal with it; instead, we would expect
him to keep turning the situation over in his mind, trying to find a solu-
tion—any solution—to his problem. This makes sense if we suppose that
anxiety is a response to a threat, but not necessarily an imminent one. Anxi-
ety keeps Harry’s attention focused on the threat, but it does not ensure
that he deals with it immediately; as a result, it does not need to focus his
attention on a limited set of responses.
Again, it might be argued that not all types of emotional response
function to prompt actions that are designed to resolve the situation. For
example, in the case of ‘backward-looking’ emotions, such as sorrow or
happiness, there is now nothing that the subject can do to influence what
has happened. A similar claim might be made about certain empathic
emotional responses—for example, fearing for a tightrope walker who is
tottering above a 30 foot drop. Indeed, it is not immediately obvious what
we should say about the function of responses of these kinds.
These issues require further discussion. For present purposes, what I
would like to suggest is that the account works well for certain kinds of
emotional appraisal, including paradigmatic cases of anger and panicky
fear. It makes sense of the urgency of these responses, their tendency to
dominate our attention, and the fact that they are typically associated with
particular kinds of action. In what follows, I shall use this account to under-
write some suggestions about the content of AP ; as I go along, I shall suggest
some ways in which the content of AP might be contrasted with the content
of an evaluative judgement about danger.

5. KINDS OF INTENTIONAL CONTENT

Intentional states can be divided into different categories. First, there are
states that register that a certain situation obtains or that a certain type of
event has occurred. These states have descriptive content, identical with the
information that they normally carry. Factual judgements possess purely
descriptive content: they function to convey information, which can be
used to satisfy a range of different goals. I have argued elsewhere that the
states produced by simple signalling systems also have purely descriptive
content. The function of a simple signalling system is to trigger some ste-
reotypical response whenever a certain condition arises. The sensory sys-
tem that triggers the eye blink reflex is a system of this kind: this system
216 Carolyn Price

works by triggering a blink whenever it senses that an object is approaching

the eye. To do this, it does not need to represent blinking or protecting the
eye as a goal; it just needs to indicate that something is approaching the eye
(Price 2001: 138-41).⁷
Secondly, some intentional states represent a goal to be achieved or an
activity to be performed. These states have directive content, determined by
the goal or activity that they normally prompt. Desires and plans have con-
tent of this kind. States with directive content differ from simple signals in
that they are normally used in some process of practical inference: in a pro-
cess of this kind, the system decides how to achieve a certain outcome, given
the information at its disposal. This might involve something as sophistic-
ated as practical reasoning, but this need not be the case: for example, a bee
might possess a navigational system that enables it to calculate how to get to
a nectar source, given its current location. For this kind of process to occur,
the system needs to represent ‘reaching the nectar source’ as a goal that the
bee’s behaviour is supposed to realize.
Desires can be ascribed purely directive content. This is because a desire
typically represents a goal that could be worth pursuing in a range of dif-
ferent contexts. As Millikan points out, however, some intentional states
represent goals that are supposed to be satisfied only when a certain con-
dition arises: for example, a bee dance will normally prompt the watching
bees to fly off to a certain location only if there is nectar at that location.
States of this kind possess a combination of descriptive and directive con-
tent: a bee dance both conveys the information that there is a source of
nectar at a certain location, and tells the watching bees to fly off to that
location (Millikan 1984: 100; 1995a:191).
In order to ascribe content to AP , we need to begin by deciding what form
of content it possesses: does it possess purely descriptive content, purely dir-
ective content, or a combination of descriptive and directive content?

6. THE FORM OF AN EMOTIONAL APPRAISAL

Many recent discussions treat emotional appraisals as judgements or

thoughts, or as in some way analogous to perceptual states, implying
that they have descriptive content of some kind. This seems correct:
an emotional appraisal is supposed to be produced in a particular type
of situation, and it is supposed to ensure that the subject’s response is
appropriate to a situation of that type. For example, a panicky appraisal
is supposed to be produced in the presence of a threat of some kind. If so,

⁷ Contrast Millikan (1984: 116–18; 1995a).

 The Content of an Emotional Appraisal 217

we should treat AP as representing the presence of a threat. For example, we

might begin by assigning it something like the following content: that cat is
very threatening.⁸
I shall return to this ascription of descriptive content later on. At this
point, I would like to turn to a different question: Should AP be ascribed
a directive content? The idea that emotional appraisals include a directive
element is not new. For example, de Sousa suggests that the content of an
emotional appraisal includes an aim or goal (de Sousa 1987: 120–1). Some
cognitive theorists have suggested that all or some emotions involve desires
(Marks 1982; Lyons 1993: 63–4). Do the considerations set out in the last
two sections support the idea that AP represents a goal of some kind?
It will be helpful, first, to make clear which aspects of the account do
not imply that AP has a directive content. First, the claim that AP functions
to trigger certain involuntary physiological and behavioural changes is per-
fectly compatible with the claim that its content is purely descriptive. Harry
does not decide how to produce these changes, and so there is no need to
suppose that he represents them as goals. In this respect, AP is operating
as a simple signal, triggering a set of stereotypical responses—just like the
sensory signal that triggers the eye blink.
We might also consider the way in which AP will normally influence
Harry’s decision-making. According to the account suggested above, AP
functions to ensure that Harry deals with the danger as a matter of urgency;
further, it functions to focus his attention on relevant information—for
example, the location of the threat. But Harry does not decide how to pri-
oritise the situation, or how to focus his attention on relevant information.
And so we do not have to suppose that he needs to represent these acts as
goals. It is enough that AP conveys the information that he has encountered
a threat, thereby triggering these motivational and cognitive changes.
However, I suggested earlier that emotional appraisals function to ensure
that the subject produces an effective response to the situation. For example,
we might suppose that a function of AP is to ensure that Harry avoids being
injured by Fluffy. Avoiding being injured is not a stereotypical response:
Harry needs to decide how to achieve this outcome, given the information

⁸ It would be possible to question whether this ascription succeeds in capturing the

full descriptive content of AP . In particular, it might be suggested that AP is supposed
to reflect, not only how things are in the subject’s environment, but also Harry’s desires
(Robinson 1983; Roberts 1988). If this were correct, it would follow that AP will
normally carry the information that Harry desires to avoid injury: in other words, the
descriptive content of AP should be expressed as: that cat is very threatening; and I desire
not to be harmed. In contrast, Michael Stocker argues that our emotions are independent
of what we desire (Stocker 1987: 64). If Stocker is right, we do not need to ascribe this
more complex descriptive content to AP . I shall not pursue this matter here, because it
turns on issues about the normal origin, rather than the function, of AP . But see n. 9.
218 Carolyn Price

that he has about his situation. For example, he might take account of
information about the proximity of the threat, the availability of exits or
suitable hiding places, and so on. If so, AP will represent avoiding injury as a
goal to be achieved.
Moreover, I also suggested that some emotional appraisals have a fur-
ther function—to direct the subject’s attention to a specific set of possible
actions. I suggested that this was true of panicky appraisals like the one
produced by Harry. The options that occur to Harry might include run-
ning away, lashing out at the threat, or hiding from it. Again, these are
not stereotypical responses: they are performed in a way that takes account
of information about the situation: for example, information about the
location of the threat, the position of an exit, and so on. This implies
that AP will represent not only the overall goal of avoiding injury, but
also a set of sub-goals: running away, lashing out at the threat, hiding, or
whatever.
Finally, this directive content will include a temporal element: it will
represent the time at which the subject is to respond to the situation. Pre-
sumably, a panicky appraisal will normally prompt the subject to act imme-
diately, but this may not be true of all types of emotional appraisal: there
may be some kinds of emotional appraisal that do not normally prompt
the subject to act immediately, but only at some time in the future. We
have already seen, for example, that anxious appraisals do not motivate
an immediate response. This highlights the need to distinguish between
the claim that an emotional appraisal functions to prompt the subject to
act immediately and the claim that it functions to ensure that the subject
prioritises the problem. An appraisal that prompts the subject to respond
at some time in the future may nonetheless function to ensure that the
subject begins to search for an effective response straight away. But this
‘straight away’ does not need to be represented, because, as we have seen,
prioritising the problem is not something that the subject needs to decide
how to do.
If all this is correct, it suggests that the content of AP should be expressed
in something like the following way: that cat is very threatening; avoid being
injured by it—by running away from it now or by lashing out at it now or by
hiding from it now.
What justifies this ascription of content to AP ? If we accept HCT, this
ascription will be justified by the history of this type of appraisal. If the
capacity to produce panicky appraisals is inherited, the crucial point will
be that this capacity sometimes enabled Harry’s ancestors to avoid being
injured by prompting them to run away from the threat, or to lash out at
it, or to hide from it. If we assume that Harry’s capacity to produce pan-
icky appraisals is learned, then this ascription will depend on the claim that
 The Content of an Emotional Appraisal 219

this capacity has, in the past, enabled Harry (or perhaps others in his com-
munity) to avoid being injured by behaving in one of these ways. In either
case, the content of AP will reflect the successes of the past.

7. THE FORM OF AN EVALUATIVE JUDGEMENT

In the last section, I suggested that AP possesses a combination of direct-

ive and descriptive content. In this section, I shall consider what form
of content we should ascribe to an evaluative judgement, for example the
judgement ‘That cat is dangerous’. In order to determine the content of the
evaluative judgement, we need to investigate its function. In other words,
we need to specify how it normally contributes to processes of theoretical
and practical reasoning.
First, an evaluative judgement normally conveys some factual inform-
ation. For example, the judgement ‘That cat is very dangerous’ will be
appropriate only if there is a substantial possibility that the cat will cause
serious harm to the subject. However, this evaluation differs from the fac-
tual judgement ‘That cat is likely to cause serious harm to me’. There is
room for different views about the root of this distinction.⁹ But it is cer-
tainly plausible to suppose that one difference between these two kinds of
state is that an evaluation normally motivates the subject to respond to the
situation in a certain way. For example, if Harry were to judge that Fluffy
is very dangerous, we would expect him to be motivated to avoid her, or at
least to treat her with caution. If this is correct, evaluative judgements, like
emotional appraisals, will possess a combination of descriptive and directive
content.¹⁰
We could accept this suggestion without supposing that emotional ap-
praisals and evaluative judgements perform their motivational function in
just the same way. I suggested earlier that emotional appraisals function
to prioritise certain considerations at the expense of others. There is no
suggestion that evaluative judgements have this effect. A general-purpose
reasoning system will normally operate by balancing the subject’s evalu-
ations and desires against each other in coming to a decision about what to

⁹ One possibility is that an evaluative judgement depends, not only on how things
are in the environment, but also on the subject’s desires or preferences. For example, it
might be suggested that an evaluative judgement about danger will normally carry the
information that the subject desires to avoid injury. (I find this suggestion attractive, but
I am not clear whether it is correct.) If it turned out that evaluations, but not emotional
appraisals, normally carry information about the subject’s desires, this would constitute
a difference in content between the two states—a contrast explained, not by a difference
in function, but by a difference in the way in which these states are normally produced.
¹⁰ Millikan tentatively endorses this suggestion (Millikan 1995a).
220 Carolyn Price

do. As a result, the judgement that Fluffy is dangerous presents the threat
posed by Fluffy as one consideration among others. Unlike an emotional
appraisal, it might well be overridden in favour of some other goal.
It is possible to make a connection between this point and claims made
by other writers concerning the distinction between evaluative judgements
and emotional appraisals. Both Michael Stocker and Peter Goldie have
suggested that the contrast between emotional appraisals and evaluative
judgements is not a matter of their content—at least, not in the sense in
which I am using the term here—but rather of the way in which their
content is entertained by or presented to the subject. Stocker suggests that
there is a distinctive sense in which the content of an emotional apprais-
al is ‘taken seriously’ by the subject (Stocker 1987). Goldie suggests that
the difference between an emotional appraisal and an evaluative judgement
is analogous to the difference between an indexical and a non-indexical
thought: in feeling fear, he suggests, ‘you are emotionally engaged with
the world, and typically you are poised for action in a new way’ (Goldie
2000: 61). Similarly, I have suggested that one difference between AP and
an evaluative judgement is that the content of AP is presented to Harry
as an immediate and overriding priority. As we have seen, however, this is
not a difference in content: because AP does not represent prioritising the
situation as a goal.
However, there are a number of ways which the content of AP might
be thought to differ from the content of an evaluative judgement. In what
follows, I shall investigate some of these differences. I shall begin by consid-
ering the directive content of the two kinds of state.

8. SOUNDING THE RETREAT

There is at least one respect in which the directive content of AP appears

to differ from the directive content of an evaluative judgement. I have
suggested that AP might be ascribed directive content taking roughly the
following form: avoid being injured by that cat—by running away from it
now or by lashing out at it now or by hiding from it now.
In contrast, there is no reason to make such a complex ascription of dir-
ective content to an evaluative judgement. This is because an evaluative
judgement does not have the function to prompt one of a specific set of
possible actions. To judge that a cat is dangerous is not specifically to be
motivated to flee from it or to lash out at it, but only to find some way to
avoid being injured by it. This evaluation will help to generate an action
only by becoming involved in a process of practical reasoning, in which the
subject’s other beliefs are brought to bear on the problem. As a result, the
 The Content of an Emotional Appraisal 221

way in which a dispassionate subject would respond to the situation will

depend on what they believe: if they believe that the best way to deal with
an angry cat is to talk gently to it, then this is what they would be likely to
do. The evaluative judgement does not embody the lessons of the past in
the way that an emotional appraisal does.
Moreover, there is no reason to suppose that the evaluative judgement
functions to prompt the subject to act immediately, rather than at some
suitable time in the future. This is a corollary of the fact that the eval-
uation does not prompt the subject to produce one of a specific set of
actions. A subject who calmly judges that they are in the presence of a dan-
gerous cat might not react immediately: for example, they might decide
to wait for something to happen (for the cat to spring, say) before taking
action.
All this suggests that the directive content of an evaluation will be sim-
pler than the directive content possessed by AP . Consider the evaluative
judgement ‘That cat is very dangerous’. The directive content of this judge-
ment might be expressed as: avoid being injured by that cat.
This suggests another way in which it would be possible to understand
Goldie’s claim that the feeling of fear leaves the subject ‘poised for action
in a new way’. The evaluative judgement ‘That cat is very dangerous’ will
normally motivate the subject to take steps to avoid being injured by the
cat; but a panicky appraisal will motivate the subject to produce one of a
specific set of responses; moreover, the appraisal motivates the subject to
respond in one of these ways right now.

9. WHAT CONSTITUTES A THREAT?

In what remains of the chapter, I shall concentrate on the descriptive con-

tent of AP .
Earlier, I suggested that the descriptive content of AP might be expressed
as: that cat is very threatening. This might prompt us to ask what exactly
constitutes a threat: is it the threat of any kind of physical injury to the sub-
ject, or is it restricted to only a subset of those? Might it include threats of
other kinds—for example, the threat of losing something that the subject
values? According to HCT, the answer to this question will depend on how
the capacity to produce panicky appraisals has benefited Harry or Harry’s
ancestors. It is possible that this explanation is very complicated, and takes
in a number of scenarios. If so, the content of AP will relate to some highly
complex relational property, or perhaps to a disjunction of such properties.
There is no reason to assume that Harry’s evaluative concept ‘dangerous’
picks out exactly the same set of threats.
222 Carolyn Price

Indeed, there is at least one reason to deny that this will be the case.
According to HCT, as we have seen, the content of AP will be determined
by the nature of situations in which the production of a panicky appraisal
has benefited Harry or Harry’s ancestors. This implies that these situations
had a feature in common: that is, they all involved a threat that could be
avoided or overcome in some way. More precisely, they all involved a threat
that could be avoided by running away, or by lashing out, or by hiding.
Where a situation involves a threat that cannot (in principle) be avoided in
one of these ways, the capacity to produce a panicky appraisal could not nor-
mally benefit the subject. This implies that AP will represent the presence,
not merely of a threat, but of a threat that can, in principle, be avoided in
one of these ways. If Harry is confronted by a threat of some other kind—
for example, if he hears that he is suffering from a dangerous illness—it
would not be appropriate for him to produce an appraisal of this kind.¹¹
At first glance, it might seem that a similar restriction will apply to
Harry’s evaluative judgements. For example, it might be thought that a
subject could not normally benefit from a capacity to evaluate a situation as
dangerous if that situation cannot be avoided. If this were correct, it would
imply that Harry’s evaluative concept ‘dangerous’ will apply only to dangers
that he can, in principle, avoid. Of course, even if this were correct, it would
imply that the descriptive content of this evaluative judgement is broader
than the descriptive content of a panicky appraisal. This is because the
directive content of the evaluative judgement ‘That cat is dangerous’ does
not concern a specific set of possible actions, but only concerns some more
general goal, such as avoiding injury. As a result, there would be no need to
suppose that the evaluative concept ‘dangerous’ concerns only dangers that
can be escaped by running away, lashing out, or hiding. So the evaluative
judgement would apply to a wider range of cases than AP .
However, I have argued elsewhere that, in the case of subjects who are
capable of a certain rather sophisticated form of reasoning, the content
of judgements and desires is not subject to this kind of restriction (Price
2001: 241–50).¹² Subjects who are capable of this kind of reasoning are
able to recognise that certain goals are (in principle) beyond their reach,
and to understand this as resulting from a combination of their own lim-
ited capacities and of some independent feature of the situation—the dis-
tance of a place, say, or the complexity of a problem. In what follows,
I shall refer to this form of reasoning as R-reasoning. A subject who is
capable of R-reasoning might sometimes save themselves time and trouble

¹¹ Anxiety might be an appropriate response, however.

¹² For other discussions of this issue, see Peacocke (1992: 129–32); Millikan (1995b);
Papineau (1996).
 The Content of an Emotional Appraisal 223

by abandoning the pursuit of an unattainable goal; or they might take

steps to prevent a certain outcome from becoming unattainable. In order
to obtain these benefits, the subject must be able to represent certain goals
as unattainable: they must be able to think of certain places as unreachable,
certain problems as insoluble, certain dangers as unavoidable, and so on. If
Harry is able to engage in R-reasoning, he might well possess an evaluative
concept ‘dangerous’ that applies to all kinds of threat, including threats that
he cannot, in principle, avoid.
Might similar considerations be made to apply to AP ? They could be
made to apply if it could be shown that Harry’s ability to engage in R-
reasoning sometimes depends on his capacity to produce panicky apprais-
als. In particular, it might be suggested that, on some occasions, a panicky
appraisal will normally prompt the evaluative judgement that the subject is
in a dangerous situation. If this were so, the subject could sometimes benefit
by producing a panicky appraisal in response to an unavoidable threat; this
is because the appraisal might sometimes prompt a judgement that was later
used in a process of R-reasoning.
However, it is far from clear that this kind of inferential connection
normally exists. Indeed, there are reasons to think that the evaluative judge-
ments and emotional appraisals are normally generated independently of
each other (Greenspan 1980; Griffiths 1990).¹³ Unless this view turns out
to be mistaken, we should conclude that there is an important distinction
between the descriptive content of AP and the descriptive content of an
evaluative judgement. The content of AP will concern only threats that
the subject is able to defuse by performing one of a limited set of actions.
Whether a threat falls into this category will depend, in part, on the prac-
tical capacities of the subject. In contrast, the content of the evaluative
judgement does not depend on the practical capacities of the subject. As
a result, the content of the evaluative judgement will be marked by a kind of
objectivity that we cannot attribute to AP .

10. THE TIMING OF THE THREAT

Philip Percival has pointed out an intriguing distinction between emotion-

al appraisals and dispassionate evaluative judgements (Percival 1992; see

¹³ The fact that a frightened person is more likely to judge that they are in a dangerous
situation is not in itself evidence for the existence of this inferential connection. This
phenomenon might arise from the fact that the feeling of fear focuses the subject’s
attention on the frightening aspects of the situation, and so on the evidence that supports
that evaluation.
224 Carolyn Price

also Maclaurin and Dyke 2002). If someone has suffered a loss, we would
expect them to feel sad; but, in most cases, we would expect their sorrow
to diminish after a time. Again, if someone has been treated very unjustly,
we would expect them to feel angry; but we would expect their anger to
fade over time. Something similar applies to fear: we would expect Harry’s
panic to fade away as soon as he has escaped from Fluffy. Indeed, in many
cases, we would think that there is something inappropriate or irrational
about an emotion that does not diminish in intensity over time. In contrast,
if someone judges that they have suffered a flagrant injustice or survived a
very dangerous encounter, we would not expect their evaluation of what has
happened to change as time passes: they should not judge that the injustice
was any less flagrant or the encounter any less dangerous because it occurred
a long time ago.
A consideration of the function of these two kinds of state can help
us to make sense of this distinction.¹⁴ I have suggested that the function
of a panicky appraisal is to mobilise physiological, cognitive, and behavi-
oural resources to deal with a threat that currently confronts the subject.
The ‘currently’ is important: Harry would not normally benefit from a
propensity to produce a panicky appraisal in response to the information
that some threat had occurred in the past, or that some threat will occur at
some time in the future. A panicky appraisal, then, will represent a threat
as present or as imminent; but not as past, or in the future. Of course,
Harry may remember a past encounter with Fluffy, and feel afraid. But his
appraisal will represent the threat as present or imminent, not as past or in
the more distant future.
The range of temporal content that an emotional appraisal can convey
will depend on the nature of the emotion. Panicky actions are likely to be
useful only in the midst of the crisis. In contrast, angry actions might help
the subject to deal with offences that are just about to happen, that are
happening, or that have happened in the recent past. This implies that an
angry appraisal might represent an offence as imminent, present, or in the
recent past—but not as belonging to the remote past or the distant future.
In contrast, there are no such restrictions on the temporal content of an
evaluative judgement. For example, if Harry can judge that Fluffy is dan-
gerous as he eyes her across the kitchen, he can also judge that Fluffy was
dangerous when he remembers the event the next day. It is easy to see how
the capacity to produce this past-tense evaluation might be of use to Harry
in drawing further inferences—for example, the inference that he ought to
avoid visiting people who live with cats. Again, Harry can also judge that
Fluffy will still be dangerous two years from now: this is because he can

¹⁴ Maclaurin and Dyke offer a similar explanation (Maclaurin and Dyke 2002).
 The Content of an Emotional Appraisal 225

benefit from a capacity to produce evaluations that concern situations that

lie in the distant future when he is making long-term plans.
As Percival makes clear, a similar point applies to the representation of
hypothetical events. Harry can produce evaluative judgements that con-
cern hypothetical situations, or even situations that are known not to have
occurred. In contrast, the account offered here suggests that a panicky ap-
praisal will always represent a situation as actual. This is because panicky
appraisals function to prompt an immediate response, which would be
redundant or even harmful in the face of a merely possible threat. The
situation may be different in the case of other emotions: for example, it
is possible to see how a subject might benefit from anxious exploration of
a merely possible situation, or even a situation that is known not to have
occurred.

11. DEGREES OF THREAT

As I have characterized it, AP represents not only the occurrence of a threat,

but also its degree: Harry’s wild flight suggests that he represents Fluffy not
simply as a threat but as very threatening. Again, he might judge that Fluffy
is very dangerous. However, there are some reasons to suppose that there
will be some differences in the way in which these kinds of state represent
the severity of a threat.
We would expect the severity of the threat represented by AP to be linked
with the intensity of the emotional episode that it triggers. The intensity of
an emotional episode appears to involve a number of elements. The more
intense Harry’s panic, the more intently his attention is focused on the situ-
ation; the more strongly he is motivated to deal with the threat; and the
more vigorous are the physiological changes that he undergoes. Presum-
ably, one of the functions of AP will be to ensure that the intensity of these
changes correlates with some variable factor in his situation. The intens-
ity of an episode of fear seems to reflect a number of factors—the severity
of the threatened injury, the probability that the injury will actually occur,
and its closeness in time. If so, AP will normally carry information about all
these factors.
There are two things to note about this. First, there is no reason to sup-
pose that these factors are represented by separate elements in the appraisal.
We would need to suppose this only if there was some differentiation in
the way in which Harry would normally respond to each of these factors.
Secondly, it might turn out that the relation between the degree of threat
represented by Harry’s appraisal and the extent to which each of these
factors is exemplified by the situation is not a linear one. In particular,
226 Carolyn Price

although it seems reasonable to suppose the degree of threat represented

by AP will increase with the probability that injury will occur, this will only
apply up to a certain point: once injury is certain, as we have seen, Harry
could not normally benefit by producing a panicky appraisal.
It follows from this that the variation represented by this element of the
appraisal might be quite idiosyncratic; there is no reason to assume that
it will be captured by some concept ‘degree of danger’ that features in
Harry’s evaluative judgements. For example, given that Harry is capable
of evaluating situations in the past, present, and future, we would expect
an evaluative judgement to distinguish the seriousness and probability of a
threat from its closeness in time. But, as we have seen, it is possible that AP
does not make this distinction.

12. CONCLUSION

I have suggested that one difference between AP and an evaluative judge-

ment has to do with the way in which the content of the two states is
entertained by or presented to the subject: in the case of AP , its content
is presented as an urgent priority. This is not a difference in content, but a
difference in the type of response that these two kinds of state are supposed
to provoke. In addition, I have argued that it is possible to identify at least
four significant differences between the content of AP and the content of
the evaluative judgements that Harry is able to make concerning danger.
1. The directive content of AP is more complex than the directive content
of his evaluative judgements.
2. The descriptive content of AP concerns only threats that Harry is able
to avoid by running away, lashing out, hiding, and so on. In contrast,
evaluative judgements may represent dangers that the subject cannot, in
principle, avoid.
3. Panicky appraisals can represent only threats that are present or immin-
ent. In contrast, evaluative judgements may represent dangers in the
past, or in the distant future; and they may represent dangers that are
merely possible or that have not in fact occurred.
4. AP represents degree of threat in a way that may be quite different from
the way in which an evaluative judgement represents degree of danger.
My conclusions are limited in two different ways. First, as I stated earlier,
they are conditional both on the success of HCT as a theory of intentional
content and also on the speculative account of the function of emotional
appraisals that I presented earlier. Secondly, in this chapter, I have offered
 The Content of an Emotional Appraisal 227

an account only of panicky appraisals. Although I have indicated some of

the ways in which the content of other types of emotional appraisal might
be thought to differ from the content of panicky appraisals, I have not
attempted to offer an account of the content of emotional appraisals in gen-
eral. Nevertheless, I hope that, by examining one relatively straightforward
case study, I have indicated how a teleosemantic approach might be used to
generate a credible and informative account of the content of other types of
emotional appraisal.¹⁵

REFERENCES

A, J. (1980), ‘A Constructivist View of Emotion’, in R. Plutchik and

H. Kellerman (eds.), Theories of Emotion (New York: Academic Press), 305–39.
 S, R. (1987), The Rationality of Emotion (Cambridge, Mass.: MIT Press).
E, P. (1980), ‘Biological and Cultural Contributions to Body and Facial
Movement’, in A. Rorty (ed.), Explaining Emotions (Berkeley: University of Cali-
fornia Press), 73–101.
(1992), ‘An Argument for Basic Emotions’, Cognition and Emotion, 6:
169–200.
E, D. (2002), ‘The Search Hypothesis of Emotion’, British Journal for the
Philosophy of Science, 53: 497–509.
F, J. A. (1983), The Modularity of Mind (Cambridge, Mass.: MIT Press).
G, P. (2000), The Emotions: A Philosophical Exploration (Oxford: Oxford
University Press).
G, P. (1980), ‘Ambivalence and the Logic of Emotions’, in A. Rorty
(ed.), Explaining Emotions (Berkeley: University of California Press), 223–50.
G, P. (1990), ‘Modularity, and the Psychoevolutionary Theory of Emo-
tion’, Biology and Philosophy, 5: 175–96.
(1997), What Emotions Really Are (Chicago: University of Chicago Press).
J-L, P., and O, K. (1992), ‘Basic Emotions, Rationality and
Folk Theory’, Cognition and Emotion, 6: 201–23.
L, R. S. (1991), Emotion and Adaptation (Oxford: Oxford University Press).
LD, J. E. (1998), The Emotional Brain: The Mysterious Underpinnings of Emo-
tional Life (London: Weidenfield & Nicolson).
L, W. (1993), Emotion (Aldershot: Gregg Revivals; first pub. 1980).
M, J., and D, H. (2002), ‘Thank Goodness that’s Over: The Evolu-
tionary Story’, Ratio, 15: 276–92.
M, J. (1982), ‘A Theory of Emotions’, Philosophical Studies, 42: 227–42.

¹⁵ I have read early drafts of this chapter at seminars at York University and at the
Open University. I would like to thank the participants for their thoughtful comments.
The chapter has also benefited from conversations with Philip Percival, Kevin Sludds,
and Fred Toates. I am grateful for their help.
228 Carolyn Price

M, R. (1984), Language, Thought, and Other Biological Categories (Cam-

bridge, Mass.: MIT Press).
(1986), ‘Thoughts without Laws; Cognitive Science with Content’, Philo-
sophical Review, 95: 47–80.
(1989), ‘Biosemantics’, Journal of Philosophy, 86: 281–97.
(1995a), ‘Pushmi-pullyu Representations’, in J. Tomberlin (ed.), Philosophic-
al Perspectives, ix (Atascadero, Calif.: Ridgeview Press), 185–200.
(1995b), ‘A Bet with Peacocke’, in C. Macdonald and G. Macdonald (eds.),
Philosophy of Psychology: Debates on Psychological Explanation (Oxford: Black-
well), 285–92.
N, K. (1995), ‘Misrepresenting and Malfunctioning’, Philosophical Studies,
79: 109–41.
N, M. (2001), Upheavals of Thought: The Intelligence of Emotions (Cam-
bridge: Cambridge University Press).
O, K., and J-L, P. N. (1987), ‘Towards a Cognitive Theory of
Emotions’, Cognition and Emotion, 1: 29–50.
P, D. (1996), ‘Discussion of Christopher Peacocke’s A Study of Concepts’,
Philosophy and Phenomenological Research, 56: 425–32.
P, C. (1992), A Study of Concepts (Cambridge, Mass.: MIT Press).
P, P. (1992), ‘Thanks Goodness that’s Non-Actual’, Philosophical Papers,
21: 191–213.
P, C. (2001), Functions in Mind: A Theory of Intentional Content (Oxford:
Oxford University Press).
R, R. C. (1988), ‘What an Emotion Is: A Sketch’, Philosophical Review, 97:
183–209.
R, J. M. (1983), ‘Emotion, Judgement and Desire’, Journal of Philosophy,
80: 731–40.
S, M. (1987), ‘Emotional Thoughts’, American Philosophical Quarterly, 24:
59–69.
T, F. (2002), ‘Application of a Multilevel Model of Behavioural Control to
Understanding Emotion’, Behavioural Processes, 60: 99–114.
T, J., and C, L. (1990), ‘The Past Explains the Present: Emotional
Adaptations and the Structure of Ancestral Environments’, Ethology and Sociobi-
ology, 11: 375–424.
 Index

adaptation 11, 24, 27, 29, 31, 33, 36 correspondence 35, 52, 106, 110, 122,
language-specific 31 136–37, 152
affordances 101, 185 fn.8 Cosmides, L. 23, 26
Agar, N. 168 Craik, K. 125
Aydede, M. 116 culture 26, 32, 37–8
material 24
bee dance 61, 107, 110, 216 evoked 23, 24–6, 32, 36
Bennett, J. 33 Cummins, R. 9–10, 47, 49–50, 51–3,
biological 64, 201 fn.7
design 100
environment 37 defect 28–30
kinds 119 Dennett, D. 117, 168
purposes 102–03, 105 Descartes, R. 75, 76 fn.9 (see also
utility 100, 102–03, 105 Cartesian)
Bloom, P. 24, 119, 123 de Sousa, R. 213–4, 217
Boghossian, P. 159–60 dispositions 103, 105–08, 132, 138,
Braddon-Mitchell, D. 159–60 155
relational 107
Camhi, J. 183–4 Donald, M. 25
Campbell, D. 15 Dretske, F. 8, 117, 120–21, 123,
Cartesian 81, 98 141 fn.17, 149, 151, 168, 201
coevolution 19, 32, 36–8, 158–9, fn.6
161, 164 Dunbar, R. 29
cognitive science 16, 19, 20, 42–3, Dyke, H. 224
48–50, 55, 96, 150, 167–69,
183, 189–90, 191, 205–6 Ekman, P. 210, 212
coherence 110, 139 fn.15, 180 ,183 encapsulated 16, 25, 32–4, 36,
communication 28, 30, 34, 147, 151, 211–12, 214
160, 165 emotions 208–27
compositional 10, 13, 106–08 Ewert, J.-P. 173, 175
computational 34, 59, 155, 169, 178, externalism 76
200
concept 24, 34–6, 78, 104, 187–88, feedback 32, 61–3
191–92 fitness 19, 28
empirical 108–09 Fodor, J. 25, 33, 42–3, 116–17, 146,
evaluative 221–23 168, 187–88, 190 fn.11, 201 fn.6,
mentalistic 44–6, 65, 72 211
of belief 76 frog visual system, 168–72, 174–5,
conditioning 104, 155, 172 179
instrumental 102 functional
operant 102 analysis 10
consumer 19, 49, 100, 105–08, categories 3
117–18, 150–53, 157, 159–60, indeterminacy 174 fn.5, 183 fn7.
202 property 94, 96, 122
mechanism 4–6, 49, 61, 63–4 role 18, 89, 93, 96
semantics 100 teleo- 61–2, 135
contraries 113 traits 12
230 Index

functionalist 54, 59 Kaplan, D. 148

functions Keil, F. 188
biological 1, 3–5, 9, 12, 42, 61, 75, kimu 7–9
100, 105, 107, 110, 144, 149–50, Kripke, S. 62, 106–07
157, 165, 209, 212
concept of 3, 60, 62 language
consumer 152 module 26–8
detector 153–54, 157 of thought 200, 205
effector 153–54, 157, 161, 165 parsing 33
etiological 10–3, 15–6, 73–4 proto- 26–7
Millikan 12–3 learning 15–18, 26, 31–2, 35, 63, 72,
normal 74, 157 109, 122–23, 125–27, 130, 137,
proper 13, 16, 33, 62, 84, 106 139–41, 155, 172, 197–9,
relational 13, 18, 123 203–04
systems account of 10–11 classical, 126–7, 137
instrumental 102
Gallistel, R. 52 mechanisms 15, 61, 212
Gelman, S. 188 social 14, 24 Fn.2
genes system 57, 106
homeo-box 102 trial-and-error 18
purposes of 103–05, 113 Leibniz’s Law 88, 97
selection of 12–16, 102, 104 Lettvin, J. 170
Gibson, J. 101, 146, 185, 190 Lewis, D. 163–5
Goldfarb, W. 53–4 Locke, J. 188
Goldie, P. 210, 220–1 Loewer, B. 141, 167
Grice, P. 33, 73 Lorentz, K. 173

habituation, 154–6, 161–2 Macdonald, G. 199

Hardcastle, V. 73 fn.5 Maclaurin, J. 224
Haugeland, J. 202 fn.6 magnetosomes 203–4
Hebb, D. 155 malfunction 11, 12, 21, 73, 229, 194
Hempel, C. 2 map
Hilbert, D. 159 cognitive 51, 55–9, 65
hippocampus 51, 57 inner 50, 55, 58, 61
Hubel, D. 200, 202 road 197, 201, 204, 206
strip- 56–7
indication 8, 20, 52, 72, 117, 149, Marr, D. 190
199–202, 205, 216 memory 25, 47, 49, 83, 155
inference 33, 107, 111, 189, 224 Millikan, R. 5, 6, 12–13, 16–20, 24,
practical 213, 216 33–5, 48–9, 52, 61–5, 115,
innate 23, 25, 27, 31–2, 109, 155, 117–20, 123, 137 fn.12, 151–53,
171, 173, 188–89, 212 157, 162, 168, 187, 190, 196,
language module 36 209–12, 216, 219 fn.10, 222
systems 27, 36–7 fn.12
intentional icon 49, 64 misrepresentation 4, 12, 59, 62, 73–4,
interpretation 26, 33, 44–5, 48, 50, 149, 167–8, 184
63, 189, 190 models 121–42
isomorphism 52–3, 201 module 25–8, 30–3, 35–6, 211,
214
Jablonka, E. 32 modularity 23–4, 26, 146
Jackson, F. 159–60
Jacob, P. 168 Nadel, L. 57, 59
Johnson-Laird, P. 214 natural kinds 115–42
 Index 231

naturalism 1, 2, 43, 48 mental 1, 4–6, 14, 17, 42–5, 49,

-istic 1, 3, 4, 17, 42–3, 50, 59, 60, 51–2, 56, 59, 107, 115, 121–22,
69, 71–3 142, 167, 195 fn, 200, 206
Neander, K. 196, 209 public 42, 44, 46, 48, 53
negation, 111–12 pushmi-pullyu 153–4, 162, 165
internal 112 semantic 8
Nettle, D. 29 visual 171
neuroethology, 169–186 representational
neuroscience 56–7, 199 content 71–2, 142, 172, 196, 197,
normative 12, 62, 72–4, 98, 151 201, 203
-ity 62, 73 fn.5, 121–2 model 45–50, 54–5, 59–62, 65–6
properties 3, 58, 122
Oatley, K. 214 system 16, 100, 105, 109–10, 120,
O’Keefe, J. 57, 59 186
ontogeny 12, 37 resemblance relation 45, 47–8, 51–2,
opacity 87, 97 56, 64
opaque 18, 85–90, 98 Russell, B. 83, 112 fn.7

Papineau, D. 5–6, 42, 64, 86 fn.1, selection 10–14, 16–18, 20, 29,
89–90, 92, 97, 123, 168 fn.3, 31–2, 38, 60–3, 65, 85–94, 96,
172, 196, 222 fn.12 98, 102–04, 109, 117–18,
Peacocke, C. 18, 95 fn.8, 108, 222 120–03, 157, 185–87, 196, 199,
fn.12 203–04
perceptual experience 19, 69–70, non-genetic 13–14
147–50, 154, 157, 160, 165 ontogenetic- 15
Percival, P. 223–5 secondary 15
phenomenal 69, 91 selectional
phylogeny 12, 37 facts 86, 89
Pierce, C.S. 52 process 123
Pietrowski, P. 19 properties 86, 90–4, 96
Plato 205 recruitment 117
Poirier, P. 199, 205 relevance 120
Port, R. 195 roles 18, 89–92, 96
Price, C. 107–08, 168 fn.3, 174 fn.5, theories of content 87, 89, 94, 98
186 Sellars, W. 45–6, 65
Prinz, J. 116, 119 semantic
propositional attitude 46, 69 fn.1, 76, conceptual role semantics, 115
205 consumer semantics 100
pyramidal cells 128–42 description 44, 59, 60, 66
indicator semantics 4, 6, 11, 117,
Quine, W.v.O. 155 149, 151
informational semantics 43,
reduction 3, 90, 118–20 115–16
reference 24, 64, 95 fn.9, 195, 204–05 meta-, 148–9, 151
regress 47–8, 53, 113 properties 17, 42–4, 48
reinforcement 103–04, 123 rules 105
representation (see also success semantics 5
misrepresentation) SINBAD theory, 125–42
asymmetric dependence theory Skyrms, B. 163–5
of, 116, 187 Stampe, D. 4, 201 fn.6
cognitive 4 Sterelny, K. 57, 168, 170
inner 61–2 stimulation 8, 155, 171, 175, 178
internal 56 proximal 109–11, 205
232 Index

stimulus 19, 25, 140 fn.16, 148, Tinbergen, N. 173

154–55, 161, 164, 172–78, toad visual system, 169–192
181–90, 200 Tolman, E. 55–7
poverty of 31 Tooby, J. 23, 26
stimuli 8, 57, 146, 156–57,
170–71, 181–84, 189, 197 van Gelder, R. 195
Stocker, M. 217, 220 Velleman, D. 159–60
substances, 120 verificationist 7, 187
supervene 2–3, 75
Wellman, H. 188
teleological 94, 100–01, 168, 187, 196 Wiesel, T. 200
theory of content 85–8, 90–1 Wittgenstein, L. 53, 62
telepathy 17, 33–4 Wright, L. 61