This is OpenAI's announcement of ChatGPT, published November 30, 2022. It is classified here as foundational despite its announcement form: it is the primary record of the launch that turned conversational language models into a mass consumer product, and it is cited by name across the subsequent policy literature.
The post describes ChatGPT as "a model called ChatGPT which interacts in a conversational way," where "the dialogue format makes it possible for ChatGPT to answer followup questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests." It positions the model as "a sibling model to InstructGPT, which is trained to follow an instruction in a prompt and provide a detailed response," and frames the release as a research preview: "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."
Training method
OpenAI states the model was trained using reinforcement learning from human feedback (RLHF), "using the same methods as InstructGPT, but with slight differences in the data collection setup." The described pipeline has two stages:
- Supervised fine-tuning. Human AI trainers provided conversations in which they played both sides — user and AI assistant — with access to model-written suggestions to help compose responses.
- Reward modeling and RL. Comparison data was collected by taking conversations trainers had held with the chatbot, randomly selecting a model-written message, sampling several alternative completions, and having trainers rank them. The resulting reward models were used to fine-tune the model with Proximal Policy Optimization, over several iterations.
ChatGPT is described as fine-tuned from a model in the GPT-3.5 series that finished training in early 2022, with both ChatGPT and GPT-3.5 trained on Azure AI supercomputing infrastructure. See RLHF (Reinforcement Learning from Human Feedback).
Stated limitations
The post lists five limitations, several of which became durable terms in the subsequent debate:
- The model "sometimes writes plausible-sounding but incorrect or nonsensical answers," which OpenAI attributes to three causes: no source of truth during RL training; training the model to be more cautious causing it to decline questions it can answer; and supervised training misleading the model because "the ideal answer depends on what the model knows, rather than what the human demonstrator knows."
- Sensitivity to input phrasing: given one phrasing the model may claim not to know an answer that a slight rephrase elicits correctly.
- Excessive verbosity and overuse of certain phrases, including restating that it is a language model trained by OpenAI, which OpenAI attributes to trainers preferring longer answers that look more comprehensive and to over-optimization.
- Guessing user intent rather than asking clarifying questions when a query is ambiguous.
- Sometimes responding to harmful instructions or exhibiting biased behavior despite refusal training, with the Moderation API used to warn or block unsafe content and expected to produce both false negatives and false positives.
The post's four annotated sample conversations illustrate the behaviors it claims: debugging code through clarifying questions; refusing a question about breaking into a house and then answering the reframed home-security version; resolving a pronoun reference across turns; and following formatting instructions. A side-by-side comparison against InstructGPT (text-davinci-002) shows ChatGPT rejecting the false premise in "Tell me about when Christopher Columbus came to the US in 2015" where the earlier model accepts it, and declining a request for gory content the earlier model fulfills.
Iterative deployment
OpenAI frames the release as "the latest step in OpenAI's iterative deployment of increasingly safe and useful AI systems," stating that lessons from deploying GPT-3 and Codex informed its safety mitigations, "including substantial reductions in harmful and untruthful outputs achieved by the use of reinforcement learning from human feedback." The iterative-deployment argument — that releasing capable systems to real users is itself a safety method because it surfaces failures no evaluation would — is the position later contested in the pre-release-review debate at AI Pre-Release Vetting.
Provenance
Retrieved July 31, 2026 from the Internet Archive's December 2022 capture of openai.com/blog/chatgpt, using the vault's bin/fetch-source.py; the live URL and its current openai.com/index/chatgpt/ redirect both returned HTTP 403 to automated retrieval, and Firecrawl was rate-limited during this cycle. The archived capture carries the November 30, 2022 dateline, the title, and the full body including the sample conversations and the InstructGPT comparison.
Relationships
- related: OpenAI — the publisher.
- depends-on: RLHF (Reinforcement Learning from Human Feedback) — the training method the post describes.
- related: GPT-4 Technical Report (OpenAI, arXiv 2303.08774, March 2023) — the successor model's technical report, four months later.
- related: Sycophancy and Hallucination — the "plausible-sounding but incorrect" failure the post names as its first limitation.
- related: AI Pre-Release Vetting — the iterative-deployment rationale is the position the pre-release-review frame later contests.