What is ChatGPT — and AI systems like it

In brief

ChatGPT is a computer program that generates text by predicting the next word in a sentence. It was trained by reading billions of web pages, books, and articles. Claude, Gemini, Mistral, and the rest work on the same principle. None of them “understand” what they write in the human sense of the word — but the results are often striking.

You type a question — what happens?

When you write “Explain photosynthesis to me” in ChatGPT, the program does not go looking for the answer in a database. It does not consult Wikipedia in real time. It does something very different: it generates its response word by word, choosing at each step the word most likely to continue the sentence well.

Think of a jazz musician. He does not replay a memorized piece. He improvises — but his improvisation rests on thousands of hours of practice, pieces heard, scales internalized. ChatGPT does the same with words. It has “read” so much text during training that it can produce coherent sentences on almost any topic.

The result resembles a conversation with a well-read person. But the mechanism behind it is fundamentally different from a human brain.

In short: ChatGPT has no reservoir of pre-written answers. For every question, it builds its response word by word, drawing from an ocean of probabilities shaped by its training. It’s statistical improvisation, not a lookup.

A very sophisticated parrot

The most useful analogy: think of the predictive text on your phone, the feature that suggests the next word when you type a text message. ChatGPT works on the same principle — except that instead of three basic suggestions, it chooses from hundreds of thousands of possible words, and it does so with a much broader context.

Your phone predicts the next word from the two or three preceding words. ChatGPT takes into account the entire ongoing conversation — sometimes several pages of text — to decide which word to write next.

This explains both its power and its limitations:

• Power: it produces fluid, structured, often relevant text on an enormous variety of subjects.
• Limitation: it can write perfectly worded sentences that are factually false, because its criterion is not “is this true?” but “is this word probable here?”

In short: take the predictive text on your phone, multiply its reach by 100,000, and feed it whole pages of context. It’s still next-word prediction — sophisticated, but blind to truth.

How it learned

ChatGPT was not programmed with rules like “if asked X, answer Y.” It was trained — meaning it was made to read an astronomical amount of text (books, websites, forums, scientific articles, computer code…) and the program learned, through statistics, which words frequently appear together and in what order.

Think of a child learning to speak. They don’t learn grammar from a textbook — they hear thousands of sentences and eventually produce their own. Training an LLM (language model — the technical term for these programs) works a bit like that, much faster and far more massively.

After this first phase, humans refined the responses: they rated the program’s good and bad outputs to steer it toward more useful and less problematic answers. That is why ChatGPT replies politely, structures its responses with lists, and refuses to help you build explosives.

In short: nobody wrote it rules. It was made to read the equivalent of millions of books, then humans told it “this answer is good, that one isn’t” until it got the hang of it. The “personality” of an AI assistant comes mostly from this second stage.

ChatGPT, Claude, Gemini: the same family

You have probably heard several names: ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Mistral (Mistral AI), LLaMA (Meta). These are all variations on the same principle: a program trained on text that predicts the next word.

The differences between them come from:

• Training data: each company uses a slightly different corpus.
• Model size: some models are larger (more computations per word), which generally makes them more capable — but also slower and more expensive.
• Fine-tuning: the way humans guided the program after initial training. This is what gives each its “character” — Claude tends to be careful, ChatGPT rather verbose, and so on.

But under the hood, the fundamental principle is the same. If you know how to use one, you can use the others without relearning.

What these systems do NOT do

It is equally important to understand what these programs are not:

• They are not search engines. They do not retrieve information in real time (unless a web search feature is explicitly enabled). Their responses come from what they saw during training.
• They are not reliable knowledge bases. They can invent facts, citations, bibliographic references — with disconcerting confidence.
• They are not conscious intelligences. They have no opinions, no memory from one conversation to the next (unless that feature is enabled), no understanding of the world. They manipulate words — very well — but that is all.

The analogy that works best: they are language calculators. Extremely powerful within their domain, but you should not ask them to be something they are not.

Key takeaways

• ChatGPT and its equivalents are programs that generate text by predicting the next word, trained on billions of pages.
• They do not “understand” what they write — they produce statistically probable text.
• ChatGPT, Claude, Gemini, and Mistral all operate on the same basic principle, with differences in training and fine-tuning.
• Their responses are often useful, but never guaranteed to be accurate — verifying important facts remains essential.
• They are neither search engines, nor databases, nor conscious intelligences: they are text generation tools.