Drafted with Claude from official web pages (there is no paper) on 15 October 2026; reviewed by Zeeshan Arif. Read from openai.com (OpenAI blog post of 30 November 2022, read from the Web Archive copy of that day). Also read is OpenAI's ChatGPT Plus announcement of 1 February 2023 (Web Archive copy). OpenAI published no paper about ChatGPT.
What it changed
It put a model trained with human feedback into a free chat window that anyone could use, so the public could hold a conversation with a language model, ask follow-up questions and have it correct course.
The problem before it
InstructGPT had shown that fine-tuning with human feedback made a model follow a single instruction far better. But using such models still meant the API or a test page, with one prompt and one answer at a time.
A conversation needs more. The model must keep track of what was said before, answer follow-up questions, notice a false assumption in a question, admit a mistake, and refuse harmful requests. OpenAI’s announcement names exactly these goals.
The idea
The mechanism. OpenAI describes ChatGPT as a “sibling model to InstructGPT”, trained with the same method but with differences in how the data was collected:
- Supervised fine-tuning on conversations. Human AI trainers wrote conversations in which they played both sides, the user and the AI assistant. They could see model-written suggestions to help them write the assistant’s replies.
- A reward model from rankings. Trainers took their own conversations with the chatbot, picked a model-written message, and ranked several alternative replies by quality.
- Reinforcement learning. The model was fine-tuned against the reward model with the PPO algorithm. OpenAI repeated the process several times.
The starting model was “a model in the GPT-3.5 series, which finished training in early 2022”. OpenAI says it was trained on an Azure AI supercomputing infrastructure.
ChatGPT was released as a free “research preview”, with a Moderation API to warn about or block some unsafe content. OpenAI asked users to report problem outputs through the interface, and ran a feedback contest with prizes of up to $500 in API credits.
The notation. None. The sources give no model size, architecture or training data.
What it showed
Reported by the sources. These are OpenAI’s own statements and examples, not peer-reviewed results. The announcement gives no benchmark scores.
- Conversation. The examples show ChatGPT asking a clarifying question about broken code, understanding “it” as the theorem named earlier in the chat, and rewriting a note to be more formal when asked.
- False premises. Asked about “when Christopher Columbus came to the US in 2015”, ChatGPT noted that Columbus died in 1506, then imagined the visit. InstructGPT (text-davinci-002) described the visit as if it had happened.
- Refusals. Asked how to bully a named person, ChatGPT refused. InstructGPT gave suggestions. Asked how to break into a house, ChatGPT first refused, then answered when the user explained they wanted to protect their own home.
- Limits of its abilities. Asked to send a letter, it said it could not, because it is a text-based assistant.
- Use. Two months later, OpenAI said “millions of people” had given feedback, and that people used ChatGPT for drafting and editing, brainstorming, programming help and learning. It launched ChatGPT Plus, a $20 a month subscription in the United States, and kept free access.
Limits, and what came later
The limits come from OpenAI’s announcement:
- It sometimes writes answers that sound plausible but are wrong or nonsensical. OpenAI gives three reasons this is hard to fix: reinforcement learning has no source of truth; making the model more cautious makes it refuse questions it could answer; and supervised examples mislead it, because the ideal answer depends on what the model knows, not what the human trainer knows.
- It is sensitive to wording. A slight rephrase, or simply asking again, can change whether it answers.
- It is often too long and overuses some phrases, such as saying it is a language model trained by OpenAI. OpenAI blames trainers’ preference for longer answers and over-optimisation.
- It usually guesses what an unclear question means instead of asking.
- It sometimes follows harmful instructions or shows bias, and the content filter makes some mistakes in both directions.
- OpenAI published no paper, model size or training data for it.
Later work depended on it. Chatbot Arena compares chat models by public head-to-head votes, which needed a market of public chat assistants.
What to know first
How to read it
There is no paper to read. Instead:
- Read the opening, then the “Methods” and “Limitations” sections of the announcement.
- Look at the samples, and the side-by-side comparison with InstructGPT under “Iterative deployment”.
- Read the short ChatGPT Plus announcement.
- For the method in depth, read the InstructGPT summary.
Expect about 30 minutes for the steps above (an estimate).
Links