ChatGPT

ChatGPT is a program that you can have a written conversation with. You type a question or a task into a text field, and the program writes an answer. This answer is not retrieved from a database, but newly formulated in that moment. Behind it is what is called a language model: a program that has learned how language is structured from huge amounts of text. ChatGPT was developed by the US company OpenAI and released on November 30, 2022. Within two months, an estimated 100 million people used it — faster than any online service before it.

The moment AI became everyday

The technology behind ChatGPT was not new in 2022. Similar language models had existed for years before, but only experts had access to them. What was new was the packaging: a website with an input field, no instructions, no programming knowledge required. Suddenly, anyone could try out what such a model could do.

For the tech industry, this triggered a chain reaction. Google, Microsoft, Meta, and Amazon presented their own AI products or accelerated their plans within a few months. Investors poured billions into AI companies. The chip manufacturer Nvidia, which supplies the computing cards for such models, became one of the most valuable companies in the world. When financial news talks about the “AI boom,” this is usually the starting point being referred to.

At the same time, a debate began that continues to this day. Schools and universities have to figure out how to deal with homework written by a program. Authors and publishers are disputing with OpenAI over whether their texts were used for training without permission. And companies are asking themselves which office jobs will change as a result.

Word prediction plus human feedback

At its core, ChatGPT does something simple: it calculates which word fits best next. Then it appends that word, calculates again, and works its way through the entire text this way. It learned to do this by working through huge amounts of text from the internet, books, and forums. In the process, billions of internal adjustment parameters were fine-tuned until the predictions became good.

However, a pure prediction program would be useless as a conversation partner. It would continue questions instead of answering them. That’s why a second step was added: humans rated many of the model’s answers and sorted them by quality. From these ratings, the model learned which style of answer is desired. This process is called Reinforcement Learning from Human Feedback and is the reason why ChatGPT sounds polite, structured, and usually helpful.

An important point follows directly from this: the model doesn’t verify anything. It generates what seems linguistically plausible, and that can be factually wrong. Such confidently presented errors are called hallucinations. Fabricated citations or incorrect dates are typical examples. Newer versions are therefore allowed to look things up on the web or perform calculations, but the underlying problem remains.

From homework helper to enterprise software

ChatGPT is most visible as a website and mobile app. Students have math steps explained to them, working professionals rephrase emails, programmers look for bugs in code. There is a free version with simpler models and paid subscriptions with more powerful ones. Generating images, analyzing files, and voice conversations are now also possible.

Less visible is the same core technology in other companies' products. OpenAI sells access via an interface, i.e., a technical connection that allows other programs to call up the models. That’s why the technology is embedded in customer service chats, translation tools, and office programs without “ChatGPT” being written on them. Microsoft uses it under the name Copilot in Windows and Office.

In the news, the name is often used as a blanket term for any text-based AI. This is imprecise. ChatGPT is a product made by OpenAI; competitors are called Gemini by Google, Claude by Anthropic, or Le Chat by Mistral. And GPT refers to the underlying model family, not the chat interface itself.