OpenAI releases ChatGPT as a free research preview

OpenAI had been publishing large language models for four years. GPT-3, released in 2020, was reached through an API and, at first, an application form. What arrived on 30 November 2022 was not a larger model but a different door: a web page with a text box on it.

The announcement described ChatGPT as a sibling of InstructGPT, trained to follow instructions and then tuned with reinforcement learning from human feedback. Human trainers had written both sides of sample conversations and ranked alternative replies, and the rankings were used to score the model’s answers. The post listed what the method did not fix. The model wrote plausible answers that were wrong, was sensitive to how a question was phrased, and would guess rather than ask what the user meant.

Usage was free during the research preview. Signing up needed an email address and a phone number. Within days people were posting transcripts of it writing cover letters, explaining error messages, drafting lesson plans, and inventing citations that did not exist.

Reporting followed within the week. The Guardian described the tool on 5 December as capable of writing code and solving complex problems, and cited OpenAI’s own statement that its answers could sound plausible while being wrong.

The underlying architecture was five years old and public, described in a 2017 paper from Google. What changed in November 2022 was the number of people who could reach it.

Corrections

• 2026-09-01 - This entry originally attributed a description of ChatGPT writing “essays, scripts” to unnamed researchers, and folded OpenAI’s own “plausible-sounding but wrong” caveat into that same attribution. The cited Guardian article describes it as capable of writing code and solving problems, and the “plausible but wrong” phrasing is OpenAI’s own; corrected to attribute each claim to its actual source.