OpenAI puts ChatGPT online: a free chatbot that answers in plain language

SAN FRANCISCO, NOVEMBER 2022 - On 30 November, OpenAI put a chatbot online that holds a conversation in plain language. It is called ChatGPT. A user types a question or an instruction, the programme answers, and the user can ask follow-up questions. OpenAI calls it a research preview: a public trial version that is free for now. The company wants to learn from users what the programme is good at and where it needs to improve.

From guessing words to a conversation

ChatGPT is built on a model from the GPT-3.5 series, whose training finished in early 2022. GPT-3.5 is an improvement on the language model GPT-3 from 2020. A language model is a computer program that keeps predicting the next word in a text, like the word suggestions on a phone. The difference is that it has first processed enormous amounts of text from the internet, so it can write whole paragraphs.

A model that only predicts words is not yet a good conversation partner. It continues a text, but does not automatically do what it is asked. It can also make up answers or write hurtful text. In March 2022, OpenAI researchers showed that a bigger model does not become better at this by itself. That is why ChatGPT received extra training, in three steps.

First, imitating. Trainers wrote example conversations. They played both the user and the AI assistant that gives a good answer. From these examples, the model learned what a useful answer looks like.

Then, ranking. The model wrote several answers to a single question. The trainers put them in order from best to worst. From these rankings, a second program learned to predict how good a person finds an answer. That program is, in effect, a judge.

Finally, practising. The chatbot wrote answers and received a score from the judge. It was then adjusted so that it gives answers that earn a high score more often. These steps were repeated several times. This method is called reinforcement learning from human feedback (RLHF).

OpenAI had used the same method before for InstructGPT, a version of GPT-3 that follows instructions better, with small differences in how the data was collected. In that research, people preferred the answers of an InstructGPT version with 1.3 billion parameters (the adjustable numbers in the model) over those of GPT-3 with 175 billion. GPT-3 is therefore more than a hundred times larger. The InstructGPT models also made things up less often and wrote hurtful text less often.

What still needs to improve

OpenAI itself names what the programme still finds difficult. It can give answers that sound good but are wrong. Phrasing a question slightly differently can produce a very different answer. When a question is unclear, it guesses what is meant instead of asking. It is often too wordy and repeats stock phrases, such as that it is a language model from OpenAI. Sometimes it also goes along with requests it should refuse, or shows a bias. It is precisely to improve these points that OpenAI asks users to rate answers.

Looking back from 2026

The trial was a success. According to OpenAI chief executive Sam Altman, ChatGPT had a million users after five days. In early February 2023, the Swiss bank UBS estimated that in January, two months after the launch, a hundred million people were already using it at least once a month. According to the bank, no consumer app had ever grown so fast: TikTok took about nine months. This turned a language model into something ordinary people use every day, without needing to know how it works.

OpenAI kept building. In March 2023 came GPT-4, which can also read images and went on to power the paid version of ChatGPT. In September 2024 came o1, a model that first works through a long chain of reasoning and only then answers, whereas ChatGPT in 2022 started writing straight away. Meanwhile, people rating answers and using that to steer a model remained a standard part of training chatbots. For many people, the chat window became the face of AI.