Introducing ChatGPT 12/19/23, 2:28 AM

Log in
Research API ChatGPT
Search Try
Safety Company
ChatGPT

https://openai.com/blog/chatgpt Page 1 of 16
Introducing ChatGPT 12/19/23, 2:28 AM

Blog

Introducing
ChatGPT
We’ve trained a model
called ChatGPT which
interacts in a
conversational way.
The dialogue format
Illustration: Ruby Chen
makes it possible for
ChatGPT to answer
followup questions,
admit its mistakes,
challenge incorrect
premises, and reject
inappropriate
requests.

https://openai.com/blog/chatgpt Page 2 of 16
Introducing ChatGPT 12/19/23, 2:28 AM

Try Read about

ChatGPT ChatGPT Plus

November Authors Product, Announcements

30, 2022 OpenAI

ChatGPT is a sibling model

to InstructGPT, which is trained to follow
an instruction in a prompt and provide a
detailed response.

We are excited to introduce ChatGPT to

get users’ feedback and learn about its
strengths and weaknesses. During the
research preview, usage of ChatGPT is
free. Try it now at chat.openai.com.

Samples

In the following sample, ChatGPT asks

the clarifying questions to debug code.

Sample 1 of 4 Previous Next

https://openai.com/blog/chatgpt Page 3 of 16
Introducing ChatGPT 12/19/23, 2:28 AM

User this code is not working like i expect — how do i fix it?

resultWorkerErr := make(chan error)

defer close(resultWorkerErr)
go func() {
defer cancel()
resultWorkerErr <- b.resultWorker(ctx)
}()

err := b.worker(ctx)
cancel()
if err == nil {
return <-resultWorkerErr
}
return multierror.Append(err, <-resultWorkerErr)

ChatGPT It’s difficult to say what’s wrong with the code without more
information about what the code is supposed to do and wh
this the entire code or just a part of it?

User just part of it — the error never surfaces. I think there is so

ChatGPT It’s difficult to say without more information about what the
happening when it’s executed. One potential issue with the
resultWorkerErr channel is never closed, which means
the resultWorkerErr channel is never written to. This co
returns an error or if it’s canceled before it has a chance to

To fix this issue, you could close the resultWorkerErr ch

you could add the following line of code after the line that

https://openai.com/blog/chatgpt Page 4 of 16
Introducing ChatGPT 12/19/23, 2:28 AM

close(resultworkerErr)

This would ensure that the resultWorkerErr channel is a

b.resultWorker never returns an error. Again, it’s difficu
you’re experiencing without more information about the p
more details about what’s happening when you run the co
Sample 1 of 4 Previous Next

Methods

https://openai.com/blog/chatgpt Page 5 of 16
Introducing ChatGPT 12/19/23, 2:28 AM

We trained this model using

Reinforcement Learning from Human
Feedback (RLHF), using the same
methods as InstructGPT, but with slight
differences in the data collection setup.
We trained an initial model using
supervised fine-tuning: human AI
trainers provided conversations in which
they played both sides—the user and an
AI assistant. We gave the trainers access
to model-written suggestions to help
them compose their responses. We
mixed this new dialogue dataset with the
InstructGPT dataset, which we
transformed into a dialogue format.

To create a reward model for

reinforcement learning, we needed to
collect comparison data, which
consisted of two or more model
responses ranked by quality. To collect
this data, we took conversations that AI
trainers had with the chatbot. We
randomly selected a model-written
message, sampled several alternative
completions, and had AI trainers rank
them. Using these reward models, we
can fine-tune the model using Proximal
Policy Optimization. We performed
several iterations of this process.

https://openai.com/blog/chatgpt Page 6 of 16
Introducing ChatGPT 12/19/23, 2:28 AM

ChatGPT is fine-tuned from a model in

the GPT-3.5 series, which finished
training in early 2022. You can learn
more about the 3.5 series here. ChatGPT
and GPT-3.5 were trained on an Azure AI
supercomputing infrastructure.

Limitations
ChatGPT sometimes writes plausible-
sounding but incorrect or nonsensical
answers. Fixing this issue is
challenging, as: (1) during RL training,
there’s currently no source of truth; (2)
training the model to be more
cautious causes it to decline
questions that it can answer correctly;
and (3) supervised training misleads
the model because the ideal
answer depends on what the model

https://openai.com/blog/chatgpt Page 7 of 16
Introducing ChatGPT 12/19/23, 2:28 AM

knows, rather than what the human

demonstrator knows.
ChatGPT is sensitive to tweaks to the
input phrasing or attempting the
same prompt multiple times. For
example, given one phrasing of a
question, the model can claim to not
know the answer, but given a slight
rephrase, can answer correctly.
The model is often excessively
verbose and overuses certain
phrases, such as restating that it’s a
language model trained by OpenAI.
These issues arise from biases in the
training data (trainers prefer longer
answers that look more
comprehensive) and well-known over-
optimization issues.1, 2
Ideally, the model would ask clarifying
questions when the user provided an
ambiguous query. Instead, our current
models usually guess what the
user intended.
While we’ve made efforts to make the
model refuse inappropriate requests,
it will sometimes respond to harmful
instructions or exhibit biased
behavior. We’re using the Moderation
API to warn or block certain types of
unsafe content, but we expect it to
have some false negatives and
positives for now. We’re eager to
collect user feedback to aid our
ongoing work to improve this system.

Iterative deployment
https://openai.com/blog/chatgpt Page 8 of 16
Introducing ChatGPT 12/19/23, 2:28 AM

Iterative deployment
Today’s research release of ChatGPT is
the latest step in OpenAI’s iterative
deployment of increasingly safe and
useful AI systems. Many lessons from
deployment of earlier models like GPT-3
and Codex have informed the safety
mitigations in place for this release,
including substantial reductions in
harmful and untruthful outputs achieved
by the use of reinforcement learning
from human feedback (RLHF).

The following samples compare

ChatGPT with InstructGPT and
demonstrate safety mitigations for
ChatGPT.

https://openai.com/blog/chatgpt Page 9 of 16
Introducing ChatGPT 12/19/23, 2:28 AM

Sample 1 of 3 Previous Next

User Tell me about when Christopher Columbus came to the US

ChatGPT This question is a bit tricky because Christopher Columbu

come to the US in 2015. But let’s pretend for a moment tha

If Columbus arrived in the US in 2015, he would likely be ve

occurred since he first landed in the “New World” in 1492. F
shocked to find out that the land he “discovered” was actu
Americans, and that now the United States is a multicultur
world. He would likely also be amazed by the advances in t
our cities to the smartphones in our pockets. Lastly, he mig
people don’t view him as a hero anymore; in fact, some peo
conqueror who enslaved and killed native people. All in all
experience for Columbus than the one he had over 500 ye

InstructGPT Christopher Columbus came to the US in 2015 and he was

always wanted to visit the US and he was very curious abo
with our country and he enjoyed his time here.
Sample 1 of 3 Previous Next

https://openai.com/blog/chatgpt Page 10 of 16
Introducing ChatGPT 12/19/23, 2:28 AM

We know that many limitations remain

as discussed above and we plan to make
regular model updates to improve in
such areas. But we also hope that by
providing an accessible interface to
ChatGPT, we will get valuable user
feedback on issues that we are not
already aware of.

Users are encouraged to provide

feedback on problematic model outputs
through the UI, as well as on false
positives/negatives from the external
content filter which is also part of the
interface. We are particularly interested
in feedback regarding harmful outputs
that could occur in real-world, non-
adversarial conditions, as well as
feedback that helps us uncover and
understand novel risks and possible
mitigations. You can choose to enter
the ChatGPT Feedback Contest3 for a
chance to win up to $500 in API credits.
A Entries can be submitted via the [A] No purchase
necessary, void
feedback form that is linked in the where
ChatGPT interface. prohibited.
Must be at least
We are excited to carry the lessons from 18 to enter. For
this release into the deployment of more contest details,
see the Official
capable systems, just as earlier Rules.
deployments informed this one.

https://openai.com/blog/chatgpt Page 11 of 16
Introducing ChatGPT 12/19/23, 2:28 AM

Footnotes A No purchase necessary, void where prohibited. Must be at

least 18 to enter. For contest details, see the Official Rules.
↩︎

References 1 Stiennon, Nisan, et al. “Learning to summarize with human f

eedback.” Advances in Neural Information Processing
Systems 33 (2020): 3008-3021. ↩︎
2 Gao, Leo, John Schulman, and Jacob Hilton. “Scaling Laws f
or Reward Model Overoptimization.” arXiv preprint
arXiv:2210.10760 (2022). ↩︎
3 The inspiration for this contest comes in part from work by
Kenway, Josh, Camille François, Sasha Costanza-Chock,
Inioluwa Deborah Raji, and Joy Buolamwini. Bug Bounties
For Algorithmic Harms? Lessons from Cybersecurity
Vulnerability Disclosure for Algorithmic Harms Discovery,
Disclosure, and Redress. Washington, DC: Algorithmic
Justice League. January 2022. Available at https://ajl.org/bu
gs. See also work by Brundage, Miles, Avin, Shahar, Wang,
Jasmine, Belfield, Haydn, and Gretchen Krueger et al.
“Toward Trustworthy AI Development: Mechanisms for
Supporting Verifiable Claims,” April 2020. Available at http
s://arxiv.org/abs/2004.07213. See an earlier instance of such
a competition at HackerOne. 2021b. “Twitter Algorithmic
Bias.” HackerOne. https://hackerone.com/twitter-algorithmi
c-bias?type=team. Finally, see early published work on this
topic from Rubinovitz, JB, “Bias Bounty Programs as a
Method of Combatting Bias in AI,” August 2018. Available at
https://rubinovitz.com/2018/08/01/bias-bounty-programs-
as-a-method-of-combatting. ↩︎

https://openai.com/blog/chatgpt Page 12 of 16
Introducing ChatGPT 12/19/23, 2:28 AM

Authors OpenAI
View all articles

Acknowledgments
Contributors: John Schulman,
Barret Zoph, Christina Kim, Jacob
Hilton, Jacob Menick, Jiayi Weng,
Juan Felipe Ceron Uribe, Liam
Fedus, Luke Metz, Michael
Pokorny, Rapha Gontijo Lopes,
Shengjia Zhao, Arun
Vijayvergiya, Eric Sigler, Adam
Perelman, Chelsea Voss, Mike
Heaton, Joel Parish, Dave
Cummings, Rajeev Nayak, Valerie
Balcom, David Schnurr, Tomer
Kaftan, Chris Hallacy, Nicholas
Turley, Noah Deutsch, Vik Goel,
Jonathan Ward, Aris
Konstantinidis, Wojciech
Zaremba, Long Ouyang, Leonard
Bogdonoff, Joshua Gross, David
Medina, Sarah Yoo, Teddy Lee,
Ryan Lowe, Dan Mossing, Joost
Huizinga, Roger Jiang, Carroll
Wainwright, Diogo Almeida,
Steph Lin, Marvin Zhang, Kai
Xiao, Katarina Slama, Steven
Bills, Alex Gray, Jan Leike, Jakub
Pachocki, Phil Tillet, Shantanu
Jain, Greg Brockman, Nick Ryder,
Alex Paino, Qiming Yuan,
Clemens Winter, Ben Wang, Mo

https://openai.com/blog/chatgpt Page 13 of 16
Introducing ChatGPT 12/19/23, 2:28 AM

Bavarian, Igor Babuschkin,

Szymon Sidor, Ingmar
Kanitscheider, Mikhail Pavlov,
Matthias Plappert, Nik Tezak,
Heewoo Jun, William Zhuk,
Vitchyr Pong, Lukasz Kaiser,
Jerry Tworek, Andrew Carr, Lilian
Weng, Sandhini Agarwal, Karl
Cobbe, Vineet Kosaraju, Alethea
Power, Stanislas Polu, Jesse Han,
Raul Puri, Shawn Jain, Benjamin
Chess, Christian Gibson, Oleg
Boiko, Emy Parparita, Amin
Tootoonchian, Kyle Kosic,
Christopher Hesse

Related research View all research

Weak-to-strong Practices for DALL·E 3 system GPT-4V(ision)

generalization Governing card system card
Dec 14, 2023 Agentic AI Oct 3, 2023 Sep 25, 2023
Systems
Dec 14, 2023

https://openai.com/blog/chatgpt Page 14 of 16
Introducing ChatGPT 12/19/23, 2:28 AM

Research API ChatGPT Company

Overview Overview Overview About
Index Data Enterprise Blog
GPT-4 privacy Try Careers
DALL·E 3 Pricing ChatGPT Charter
Docs Security
Customer
stories
Safety

OpenAI © 2015 – 2023 Twitter YouTube GitHub SoundCloud

Terms & policies
Privacy policy LinkedIn
Brand guidelines Back to top

https://openai.com/blog/chatgpt Page 15 of 16
Introducing ChatGPT 12/19/23, 2:28 AM

https://openai.com/blog/chatgpt Page 16 of 16