ChatGPT
OpenAI launches ChatGPT as a free research preview
View full image ↗ How was this image made?
The creative brief sent to the image model. These are instructions, not a record of what actually happened.
Original prompt & settings (JSON) ↗The launch page promised a bot that could admit its mistakes — and warned, in the same announcement, that it often produced plausible falsehoods.
30 November 2022: a free research preview
OpenAI publicly introduced ChatGPT on 30 November 2022 and invited people to try it, free, as a research preview.
The chatbot ran on a dialogue interface designed to answer follow-up questions, challenge incorrect premises, acknowledge mistakes and reject some inappropriate requests. OpenAI identified the launch model as a fine-tuned member of the GPT-3.5 series, trained on Microsoft Azure AI supercomputing infrastructure.
Trainers who played both sides of the conversation
The announcement described supervised fine-tuning in which human AI trainers wrote both halves of sample conversations — playing the user and the assistant — followed by reinforcement learning from rankings of alternative responses. OpenAI called ChatGPT a sibling of InstructGPT and said transformed InstructGPT data was mixed into the dialogue dataset.
Users were enlisted as testers from the first day. OpenAI asked for reports of problematic answers and advertised a feedback contest offering up to $500 in API credits.
What launched, and what did not
ChatGPT was the product and the interface; GPT-3.5 was the model family beneath it. GPT-4 did not reach ChatGPT Plus until March 2023, and the November release was a free, text-based preview rather than the later multimodal subscription product.
What shipped alongside the chatbot was a warning. OpenAI said in the same announcement that the system could produce plausible but false answers, guess at ambiguous requests and sometimes respond to harmful instructions.
Sources
Researched 23 Aug 2026 6 sources 2 audit passes, 1 issue open
How this was checked
Researched from the web into a fact sheet, rewritten from that sheet, then audited against it by a different model. How the pipeline works →
What the sources leave uncertain
- The date and public-release status are directly documented by OpenAI.
- The frequently repeated 100-million-user figure was a UBS estimate based on Similarweb data for January 2023, not an audited OpenAI count or a launch-day figure.