http://dx.doi.org/10.14236/ewic/EVA2019.

22

Artificial Intelligence and Problems in

Generative Art Theory

Philip Galanter
Texas A&M University
College Station, TX, USA
galanter@tamu.edu

In previous writing I’ve described what has arguably become the most widely cited theory of
generative art. Based on notions from complexity science, and in particular Murray Gell-Mann and
Seth Lloyd’s notion of “effective complexity,” I argue that generative art is not a subset of
computer art. Rather, generative art turns on the use of autonomous systems and the artist ceding
control to those systems. As part of this theory for generative art, I’ve introduced a series of
problems. These are not problems in the sense that they require single correct solutions. Rather
they are questions that the artist will consider when making a piece; that critics and historians will
typically address in their analysis; and that insightful audience members will ponder. They are
problems that typically offer multiple opportunities and possibilities. It is notable that, for the most
part, these problems equally apply to both digital and non‐digital generative art; to generative art
past, present, and (it is believed) future; and to ordered, disordered, and complex generative art. In
addition, these same problems or questions are generally trivial, irrelevant, or nonsensical when
asked in the context of non‐generative art. In a sense the applicability of these questions can
cleanly divide art into generative art and non-generative art. More importantly, the exploration of
these questions can illuminate the analysis and critique of generative art. More recently a new form
of neural-network-based artificial intelligence called “deep learning” has appeared on the scene.
Deep learning has been applied to digital art creation. In this paper I explore whether the problems
in generative art noted above hold up well in this new artificial intelligence context for generative
art. The conclusion reached is that our current complexity-based theory of generative art can
easily assimilate the use of deep learning.

Art theory. Generative art. Neural networks. Inceptionism. Deep learning. Artificial intelligence. Complexity theory.

1. INTRODUCTION Beyond this definition a number of additional ideas

were outlined. In particular, the notion of effective
In 2003 I wrote a paper that laid out the core ideas complexity was used as a framework for sorting
for a theory of generative art using notions from various generative systems. The idea is that
complexity science as a context (Galanter 2003). systems can vary in complexity, and while simple
The key element was the idea that what is definitive systems are typically either very highly ordered or
about generative art isn’t what it is, but rather how it very highly disordered, complex systems exhibit
is made. In particular, I suggested that generative both order and disorder (Gell-Mann 1995). How this
art is created when an artist cedes some degree of is applied to generative art systems is illustrated in
control to an autonomous system that creates, or figure 1. The move to a complexity-based context
is, the art. Since its publication this paper has brings with it the opportunity to leverage concepts
arguably served as the most frequently cited theory like emergence, connectionist agents, feedback,
generative art theory in the literature. The often- nonlinearity, deterministic chaos, self-organization,
quoted definition is: and so on.
Generative art refers to any art practice where
the artist uses a system, such as a set of natural
language rules, a computer program, a machine,
or other procedural invention, which is set into
motion with some degree of autonomy
contributing to or resulting in a completed work
of art.

112
 Artificial Intelligence and Problems in Generative Art Theory
Philip Galanter

weighted sums. That middle layer is then

connected to a layer of nodes that represent output
values. A neural network “learns” by processing a
body of input data and iteratively updating the
network weights. Later, when presented with novel
input, patterns in the input will result in patterns in
the output based on previous learning.

Figure 1: Generative art methods by effective

complexity.

There are additional implications offered by this

theory. For example, generative art is not a subset
of computer art, but rather it is technology agnostic.
And indeed when one considers the ancient use of
symmetry and tiling found inscribed on artefacts Figure 2: A very simple traditional neural network.
some 77,000 years old, it’s clear that generative art
is as old as art itself (Balter 2002). Early typical practice was to limit the number of
hidden layers to just one or two. Training neural
Over the years I’ve refined and expanded this view networks with more layers, typically to attack more
of generative art. This has been thoroughly difficult problems, was found to be impractical.
compiled in a recent chapter (Galanter 2016a). More layers, and more nodes, require gathering
There it is noted, for example, that generative art more training data to extract more complex
only requires a weak form of autonomy, i.e. the patterns. In addition, the processing requirement
sense that word is used in robotics. This avoids a for more nodes and more data increases
number of philosophical complications that a exponentially due to the growing combinatorics of
stronger form of autonomy would invite. the connections.

In addition, some find it surprising that not all rule- More recently “deep learning” has overcome this
based art is generative art. For example, some perceived limitation. The need for data has been
rules are suggestive but insufficient to determine a satisfied to some extent by the “big data”
final design. Others are constraint rules that tell the movement whereby large databases have been
artist what not to do, but again are not sufficient to released for public use and access via the Internet.
fix a specific form. Of particular interest to artists are image databases
of artworks that have been released by various
1.1 Artificial neural networks and deep learning museum and university initiatives. These can be
used to learn art styles, and then modify new input
In this same chapter the topic of artificial neural images to produce output images exhibiting that
networks (or simply “neural networks”) is briefly style. In addition, the need for increased
mentioned. Neural networks are inspired by computation has been met by Moore’s Law and
nature’s biological computer, the brain. Just as recently affordable GPU-based computing clusters.
neurons establish networks where associations are
created based on the strength of synapse Deep learning systems now typically implement
connections, artificial neural networks use weighted dozens of layers, and layer operations can include
virtual connections to associate various input not only traditional weighted sums, but also the
patterns with corresponding output patterns. application of convolution matrices and other
techniques. Some of the first art applications were
Traditional neural networks typically have three explored by Google Research, resulting in a style
layers. A layer of nodes for input data is widely and technique known as inceptionism (Mordvintsev
connected to a middle “hidden” layer that gathers et al. 2015).

113
 Artificial Intelligence and Problems in Generative Art Theory
Philip Galanter

2. PROBLEMS IN GENERATIVE ART THEORY human or the computer?” Many will reflexively
answer that the programmer, i.e. the human, is the
In the same chapter noted above I introduced a artist. This question is particularly sharp in that it
number of problems in generative art theory resonates with a major vector in humanities
(Galanter 2016a). These are not problems requiring discourse, that being poststructuralism and notions
single correct solutions. Rather, these are problems such as “the death of the author.” Some might see
that invite consideration by the artist when creating digital generative art as the very embodiment of the
generative art; that theorists will want to consider in shift of attention away from traditional views of the
their analysis; and that well-informed audience author.
members will want to ponder. These are problems
that encourage multiple possibilities and In other writing I’ve introduced a point of view I’ve
opportunities. called “complexism.” It is an attempt to reconcile
the decades long “war” between the modernist
It is posited that these problems generally apply to culture of science and the postmodern culture of
all forms of generative art including digital and non- the humanities (Galanter 2016b). This is done by
digital forms, as well as past, present, and future applying insights from complexity science in a
work. But of equal importance, these same cultural studies setting. In that critique I offer a
problems when applied to non-generative art can theory of authorship that suggests when authorship
seem irrelevant, trivial, or even nonsensical. would be most appropriately credited to the
computer rather than the programmer. And, in fact,
The fact that this body of problems applies to all the example offered is a subset of deep learning-
generative art, but only generative art, is a very based generative art.
good sign in terms of art theory. It means that this
generative art theory isn’t arbitrary or forced, but Whatever side one might take in the question of
rather that it has content significance and whether a computer can truly be considered an
consistency. author, it’s clear that it is a non-trivial and
meaningful question. And the introduction of deep
But how well will these problems capture new, learning AI-based generative art only increases the
previously unanticipated, forms of generative art? importance of that question.
An analysis of the new forms of artificial
intelligence-based generative art afforded by deep Compare this, for example, to the question “Who is
learning offers a strong test case. We can try to the author of the Mona Lisa?” In the case of non-
apply these problems in generative art theory to generative art the question is not a philosophical
deep learning art. If they are found to be relevant conundrum, it is merely a question of fact.
rather than irrelevant, substantial rather than trivial,
and sensible rather than not, this should reinforce 2.2 The problem of intent
overall confidence in the associated generative art
theory. The problem of intent asks, “Why is the artist
working with and ceding control to generative
In the remaining sections each of these problems is systems?”
introduced, and then discussed in the context of
generative and non-generative art. Then the It was noted earlier that the category “generative
problem is considered relative to AI-based art” is a reference to how the art was made, but it
generative art afforded by deep learning. Specific says nothing about why the artist chose to work
responses to the problem are not the point. The that way. For example, John Cage, William
question is whether these problems are sensibly Burroughs, and Ellsworth Kelly all used
applicable to deep learning-based generative art at randomization to create art. But they did so for
all. (The noted chapter presents a much more quite different reasons. Cage used randomization
detailed treatment of these problems, but without to put into practice a Zen attitude of non-judgement
the following consideration of AI-based generative in aesthetic consideration. Burroughs, on the other
art.) hand, hoped to unleash the unconscious via
randomization as a Dada-esque tactic. And Kelly’s
2.1 The problem of authorship intent was to explore the creation of form through
random erosion.
Regarding generative art, the problem of
authorship asks, “How do traditional views of For some artists it is the generative system itself
authorship shift regarding credit, expression, and that is the topic of exploration. For others it is
provenance?” merely a means to some practical end. And still
others are looking to exercise what I’ve called “truth
When first encountering digital generative art, a to process” where a visible trace of the generative
novice will frequently ask “Who is the artist, the system is evident in the artefact (Galanter 2009b).

114
 Artificial Intelligence and Problems in Generative Art Theory
Philip Galanter

 Art as open concept via family resemblance

One common motivation among artists is to use (neo- Wittgensteinianism)
generative systems to inject an element of surprise  Art as institution
in their art making practice. Perhaps someday AI-  Art as historical definition
based systems using deep learning will act as full-
fledged studio collaborators. A discussion of these theories is beyond the scope
of this paper, but at least some of them are more or
But however used, and for whatever reason, deep less what everyday language would suggest.
learning systems clearly invite the question of intent However, that’s not to imply that carefully nuanced
common to all generative art practitioners. But philosophic discussion is unnecessary (Carroll
asking why a non-generative artist uses generative 1999).
systems is clearly nonsensical.
Of these theories, the one most likely to be
2.3 The problem of uniqueness problematic with regard to authenticity, is “art as
expression.” This was an idea born of romanticism,
The problem of uniqueness asks, “Does it diminish a period when ascendant science seemed poised
the value of the art when unique objects can be to dominate thinking about the world, and art was
mass‐produced?” driven inward as a way to explore emotions and
subjectivity.
Traditionally works of art have been treasured as
unique and thus rare objects. Walter Benjamin If one’s basis for art is the expression of feelings
offered a critique of once new art technologies for and emotions, it might seem obvious that
mechanical reproduction such as printmaking and generative systems without feelings and emotions
photography. Benjamin declared such work to have cannot create art. One possible retort would be that
a diminished “aura.” Today digital media art and the feelings and emotions being explored are those
Internet distribution has allowed dematerialization of the artist, and the generative system is used as a
to the point where duplication approaches zero tool towards that end. A stronger response would
cost. reference other theories of art, and then make the
point that expression does not have a monopoly on
Digital generative art introduces a completely new art making.
paradox. Rather than offering an endless supply of
copies, it provides an endless supply of original and In the case of AI-based generative systems such
unique artefacts. The apparently oxymoronic as deep learning, this problem teases issues
phrase “mass‐produced unique objects” in fact around sentience and consciousness. I’m unaware
describes the reality of generative art. Some of artists currently claiming to have created
generative artists exercise this new paradigm so as conscious systems, but it seems to be only a
to be in itself the content of the work. matter of time before some artists will claim that
their generative system is capable of expression in
Deep learning AI-based generative art can easily a way no different than found in human artists.
generate mass-produced unique objects, and in
this respect is no different than previous forms of To underscore the obvious, non-generative art
generative art. doesn’t participate in the problem of authenticity.
Human artists are assumed to be thinking and
2.4 The problem of authenticity feeling agents.
The problem of authenticity asks, “Given that it is in 2.5 The problem of dynamics
part created by an unemotional and unthinking
system, is generative art really art at all?” The problem of dynamics asks, “Must generative
art change over time while being exhibited to an
There are those who will insist that generative art is audience?”
simply not art. A critique of this issue almost
immediately engages what many consider a This is a topic of debate among artists and critics.
foundational art theory question, “What is art?” One type of generative art is that where the
generative system is left behind in the studio, and
In the realm of analytic philosophy there are a the artwork’s form is a fixed artefact. Another type
number of theories in this regard. They include: is where the artwork is the generative system. As
such the audience experiences the work as form
 Art as representation that is not only changing over time, something that
 Art as expression even a film does, but is also not predetermined and
 Art as significant form devoid of surprise after having been viewed a first
 Art as experience time.

115
 Artificial Intelligence and Problems in Generative Art Theory
Philip Galanter

Digital generative art raises ontological issues as to

Generative art created by deep learning systems is where the art resides. But such debates are not
mostly of the first type, but real-time dynamic limited to digital generative art. For his wall
generative art is also possible given adequate drawings Sol LeWitt would author construction
computational power. directions, and then different assistants in different
locations would follow those directions to install the
In any case, by its very nature non-generative art drawing. (Such work is generative when execution
does not enter into such discussions. And AI-based of such directions has the potential of yielding non-
deep learning generative art clearly does. identical instances.) Again one can ask where the
art is. Is it in the paper instructions? Or is it in the
2.6 The problem of postmodernity text, and the paper copy is merely a representation
of the text? Or is the art, in fact, the physical
The problem of postmodernity asks, “Is generative drawing on the wall?
art an unavoidably postmodern approach to art?”
With digital generative art there is an additional
Of the problems noted here, the problem of twist. Generative artists will sometimes make their
postmodernity is perhaps the one that has a limited code available, and arbitrary people can download
life expectancy. One can reasonably wonder it, modify it, and then run it to create their own
whether the attitudes and issues around variations. This problematises the traditional role of
postmodernism will remain in current discourse in a heroic single artist creating a fixed masterpiece.
twenty or thirty years. Now where is the art? Is it the original artist’s
artefact? Is it the code? Is it the code as modified?
Similar to the problem of dynamics, the relationship Or is it the second artefact made by a second
between generative art and postmodernism is a artist?
topic of debate among artists and critics. In
generative art some see the reification of post- Reasonable people can disagree as to the answers
structural issues regarding authorship. And artificial here, but there is little doubt that there are
life inspired generative systems resonate with the questions in this realm relevant to all generative art.
concepts of simulacra and simulation offered by And indeed these same questions easily apply to
Baudrillard (1994). AI-based generative art systems. But these
considerations are irrelevant to non-generative art.
But generative art can also be viewed as a
response that reverses postmodernity’s corrosive 2.8 The problem of creativity
claims. By harnessing systems such as reaction-
diffusion, evolution, artificial life, and other natural The problem of creativity asks, “Are generative
processes, generative art can rescue art from systems creative? What is required to create a truly
postmodernism’s distain for formalism. Generative creative computer?”
art can rescue truth and beauty from the nihilistic
relativism and social construction of The philosopher Margaret Boden has written that
postmodernism. “Creativity is the ability to come up with ideas or
artefacts that are new, surprising, and valuable”
Generative deep learning AI systems, by vaguely (Boden 2004). But few would want to argue that
simulating natural neurology, participate in both of digital generative art systems, let alone non-digital
the above; they can build up or break down the systems, are capable of forming “ideas.” But
influence of postmodernism on aesthetics as noted perhaps her allowance for artefacts would allow
above. generative art systems to be considered creative.

Like generative art in general, AI-based generative A reasonable step in the direction of machine
art comfortably fits into this discourse. The vast creativity in the arts would be the capability to
majority of non-generative art, however, was discriminate between high- and low- quality art. For
practiced long before postmodernism arrived on the example, an evolutionary computing system could
scene, and is largely irrelevant relative to that use such a capability as a fitness function. This
discourse. would lead to an overall generative art system that
modifies its own behaviour.
2.7 The problem of locality, code and
malleability In previous writing I’ve taken a different tack than
Boden (Galanter 2009a). The core idea is that the
The problem of locality, code, and malleability asks, difference between a non-creative and creative
“Is the art in the object, the system, the code, or system is the difference between a “complex
something else entirely?” system” and a “complex adaptive system.” As
noted previously, complex systems are those that

116
 Artificial Intelligence and Problems in Generative Art Theory
Philip Galanter

exhibit features such as emergence, connectionist serves as a good example of the notion of “truth to
agents, feedback, nonlinearity, deterministic chaos, process” mentioned earlier.
self-organization, and so on. Complex adaptive
systems have these same features, but in addition Similar to both digital and non-digital generative
they modify their structure or behaviour to maintain systems, deep learning AI-based generative art
their integrity in response to changes in the systems can be at both extremes. They can be
environment. So, for example, a weather system is about generativity, or not about generative systems
a complex system, but a beehive is a complex at all, or something in between.
adaptive system.
However, the question of whether generative art
There is more that can be said about this paradigm can be about more than generative systems is
for creativity, but it’s worth pointing out here that trivially irrelevant when asked in the context of non-
this approach avoids troublesome considerations generative art.
regarding consciousness and awareness, and it
can be extended beyond humans.
3. CONCLUSION
When it comes to non-generative art any creativity
is credited to the artist and not the artist’s brushes The goal here was to determine whether deep
or pencils. But in the case of generative art, where learning AI-based generative art would comfortably
the system surprises the artist it’s reasonable to fit within generative art theory that is based on the
wonder whether the artist has contributed all of the artist ceding control to autonomous systems for the
creativity on display. And in the case of AI-based creation of art.
generative art systems the problem of creativity
becomes even more relevant. It was noted that a set of “problems” can be used in
the discussion of generative art, and that these
2.9 The problem of meaning same problems can seem irrelevant, trivial, or even
nonsensical when applied to non-generative art.
The problem of meaning asks, “Can and should
generative art be about more than generative So these problems can serve a dual function. On
systems?” the one hand they can invite a discussion for any
given piece of generative art. In addition, the
One of the advantages of theorizing generative art degree to which the problems are relevant to a
as simply a way of making art is that it maximises given kind of art production method can indicate
artistic options. As noted in the problem of intent, how appropriate it is to think of that method as a
the same generative system can have different generative art system.
meanings to different artists.
The ease with which those problems can be
For some a generative system might simply be a applied to deep learning AI-based artworks is
pragmatic solution to a production need. For strong confirmation that they fit within this paradigm
example, in animated filmmaking one might use an for generative art. And in the analysis of deep
L-system-based generative tool for populating a earning Ai-based generative art no need to modify
forest scene with trees. This would be much easier, this current generative art theory was discovered
i.e. less expensive, than modelling hundreds of along the way.
trees by hand. But the film itself would not be about
generative art, or L-systems, or even trees.
4. REFERENCES
However, some generative artworks are about
Balter, M. (2002) From a Modern Human's Brow –
generative systems and little more. For example,
or Doodling? Science, 295, pp.47–248.
Haacke’s Condensation Cube is a non-digital
generative artwork. It is a clear, sealed, Baudrillard, J. (1994) Simulacra and simulation.
approximately 76 cm cube, with about a quarter of Ann Arbor, University of Michigan Press.
an inch of water in it. The water evaporates and
Boden, M. A. (2004) The creative mind: myths and
then condenses on the walls of the cube. This
mechanisms. London, Routledge.
creates ever-changing patterns of condensation
and droplets that flow back to the bottom of the Carroll, N. (1999) Philosophy of art: a
cube. The piece isn’t merely about the patterns on contemporary introduction. London, Routledge.
the cube. It’s actually about the generative system
creating the patterns. Condensation Cube was Galanter, P. (2003) What is Generative Art?
originally titled Weather Cube, but Haacke changed Complexity theory as a context for art theory.
International Conference on Generative Art, Milan,
the name to better reflect its literal function. It
Italy, 2003, Generative Design Lab, Milan
Polytechnic Art.

117
 Artificial Intelligence and Problems in Generative Art Theory
Philip Galanter

Galanter, P. (2009a) Thoughts on Computational Galanter, P. (2016b) An introduction to

Creativity. Computational Creativity: An complexism. Technoetic Arts: A Journal of
Interdisciplinary Approach, Dagstuhl, Germany, Speculative Research, 14(1–2), pp.9-31.
2009, Schloss Dagstuhl – Leibniz-Zentrum fuer
Gell-Mann, M. (1995) What is complexity?
Informatik, Germany.
Complexity – John Whiley and Sons, 1(1), pp. 16-
Galanter, P. (2009b) Truth to Process – 19.
Evolutionary Art and the Aesthetics of Dynamism.
Mordvintsev, A., Olah, C. & Tyka, M. (2015)
International Conference on Generative Art, Milan,
Inceptionism: Going Deeper into Neural Networks.
Italy, 2009, Generative Design Lab, Milan
http://googleresearch.blogspot.com/2015/06/incepti
Polytechnic Art.
onism-going-deeper-into-neural.html (retreived
Galanter, P. (2016a) Generative Art Theory. In November 11, 2015).
Paul, C. (ed.) A Companion to Digital Art, John
Wiley & Sons, Hoboken.

118