How does ChatGPT work? A simple explanation

A language model, not an encyclopedia

ChatGPT is a service you use in a web browser or as an app. You type a question or a request, and ChatGPT replies. OpenAI lists uses such as summarizing and organizing information, translation, coding help, research and creative ideas.

At its heart is a so-called language model. It is not a reference book with ready-made answers. It is a mathematical model that has learned patterns in language. One key point: in a normal reply, ChatGPT does not look up whether something is true. It produces text that fits the question.

Germany's consumer advice centres therefore describe chatbots as a complement to search engines. A search engine shows links to pages. A chatbot writes an answer itself.

How ChatGPT learned

According to OpenAI, such a model is built in several stages. First, training data is collected. OpenAI names three sources: freely available content from the internet, data from partner companies, and content provided by users, human trainers and researchers. Increasingly, artificially generated training examples are added as well.

In the so-called pre-training stage, the model practises filling gaps in sentences, that is, predicting the next word. At first its guesses are close to random. After a vast number of examples it recognizes patterns and picks the right word more and more often. This is followed by post-training, in which the model learns, among other things, to answer helpfully and safely.

What is stored is not text but numbers. Experts call them weights or parameters. During learning, these numbers are adjusted slightly again and again. OpenAI stresses that the model does not keep copies of its training texts. That is the provider's account. At the same time, OpenAI says it removes duplicate content to reduce the risk of repeating training data. So the risk apparently cannot be ruled out entirely.

How a reply is produced

When you ask a question, ChatGPT breaks the text into small units called tokens. A token can be a whole word, part of a word or a punctuation mark. The model then calculates which unit fits best next, adds it, and repeats this until the reply is complete.

Often several continuations would fit. OpenAI explains that this is why an element of randomness is built in. The same question can therefore produce different answers.

What the model learned in training has a cut-off point. For current questions, ChatGPT can also search the web. According to OpenAI, it writes one or more search queries, sends some of them to partner search providers, reads the results and writes an answer with source citations. As of October 2026, web search is available on all plans, including the free one.

Why ChatGPT can be wrong

Sometimes ChatGPT makes things up that sound convincing but are false. This is called hallucination. OpenAI explains it like this: the model learns from text that does not say what is true and what is false. Spelling and grammar follow consistent patterns, so the model gets them right. Rare individual facts, such as the birthday of a little-known person, cannot be derived from patterns.

OpenAI adds that many performance tests reward guessing and penalize an honest I don't know. According to the company, newer models hallucinate less often, but the problem is not solved.

Germany's consumer advice centres point out that answers are not checked for accuracy and are no substitute for professional advice. Answers based on web search can be wrong too. OpenAI itself advises opening the cited sources and checking whether they really support the statement.

What happens to what you type

For its consumer services, OpenAI may use your conversations to train its models further. You can turn this off. In the settings, under data controls, there is an option to improve the model for everyone. According to OpenAI, once it is off, new conversations are no longer used for training. One catch: if you rate a reply with thumbs up or down, OpenAI says the entire conversation may be used for training.

According to OpenAI, temporary chats are not used for training and do not appear in your history. On business and education plans, OpenAI does not train on inputs by default.

For training with data from Europe, OpenAI relies on legitimate interest under the EU's General Data Protection Regulation. Whether that holds in every case is a legal question we do not assess here. Consumer advocates in any case advise against entering sensitive personal information.

More than a chat

As of October 2026, ChatGPT has two main areas. Chat is meant for quick questions and conversations. Work, according to OpenAI, is a so-called agent for longer, multi-step tasks, such as research or creating a document. Work is only available on paid plans. For programming there is also Codex.

The foundation stays the same. Work is also based on a language model that produces text word by word, except that it can additionally use tools such as a browser. The models themselves change often, and their names and capabilities shift within a few months. The underlying principle does not.