OpenAI introduces the ChatGPT research preview

Background

Before ChatGPT, most people met large language models through developer tools or demonstrations rather than a chat window they could type into directly. OpenAI had earlier introduced InstructGPT, a model trained with reinforcement learning from human feedback (RLHF). OpenAI later described ChatGPT as a sibling model to InstructGPT, trained with the same methods but with slight differences in data collection.^[1]

What happened

On 30 November 2022 OpenAI introduced ChatGPT, describing it as a model that interacts conversationally, and made it free to use during a research preview.^[1] The company said the dialogue format lets the model answer follow-up questions, admit its mistakes, challenge incorrect premises and reject inappropriate requests.^[1]

On 1 February 2023 OpenAI launched a pilot paid subscription plan, ChatGPT Plus.^[2]

Key details

Training combined several steps. Human AI trainers wrote conversations in which they played both the user and the AI assistant. This dialogue data was mixed with the InstructGPT dataset. The trainers then ranked alternative model responses, which was used to build a reward model, a model that scores answers according to human preferences. Finally, the model was fine-tuned with Proximal Policy Optimization (PPO), a reinforcement learning algorithm, over several iterations.^[1]

ChatGPT was fine-tuned from a model in the GPT-3.5 series that finished training in early 2022. ChatGPT and GPT-3.5 were both trained on Azure AI supercomputing infrastructure.^[1]

OpenAI listed several limitations at launch: answers that sound plausible but are incorrect, sensitivity to how a question is phrased, excessive verbosity, guessing at ambiguous questions, and sometimes responding to harmful instructions. It used its Moderation API to warn about or block some unsafe content.^[1]

On 6 November 2023, at its first DevDay, OpenAI said ChatGPT had about 100 million weekly active users and that more than 92% of Fortune 500 companies used the platform.^[3] CNBC described ChatGPT at the time as having broken records as the fastest-growing consumer app in history.^[3]

By August 2025 OpenAI said it expected ChatGPT to reach 700 million weekly active users.^[4] Weekly active users counts the distinct people who use a service in a typical seven-day period, so the figure indicates reach rather than how deeply people use it.

Impact

By presenting the model as a chat window that anyone could use, ChatGPT brought large language models to a far wider audience than developer tools had. Someone could now test a language model's strengths and weaknesses in a single conversation, without writing any code. Its listed limitations, especially answers that are plausible but wrong, became a recurring subject for users and researchers, and checking model output remains an active area of work. The free research preview followed by a paid subscription also set a pattern of how such products would be offered: broad access first, with a paid tier layered on top.