Generative AI: how does it work?
You've probably heard of ChatGPT, seen computer-generated images go by, or asked an app to draft an email for you. Behind all of this, one phrase keeps coming up: generative AI. But what's actually hiding behind that slightly intimidating term? Good news: you can understand it perfectly well without being a computer scientist.
What does "generative" mean?
For a long time, computers were mostly good at sorting and recognising. You'd show them a photo and they'd answer "that's a cat" or "that's a dog". Useful, but limited: the machine simply filed things into boxes.
Generative AI does something new: instead of merely recognising, it creates. It can write a text that didn't exist before, draw an original image, compose music, or suggest computer code. The word "generative" comes precisely from that: it generates, it produces content.
A simple intuition: a very, very gifted autocomplete
To understand how it writes text, think of your phone's keyboard. When you type "I'm heading home to the…", it suggests "house". It has noticed that after those words, "house" is a likely choice.
Generative AI does the same thing, but on an incomparable scale. Instead of suggesting a single word, it strings together whole sentences, paragraphs, pages. At each step it asks itself one question: "given everything that came before, what's the most likely word now?" Then it starts again, word after word, until it has formed a complete answer.
This is a deliberately simplified picture, but it captures the essence: the AI isn't reciting a text stored away in a drawer somewhere. It rebuilds it on the fly, drawing on patterns of language it has learned.
How does it learn?
Before it can answer you, the AI goes through a long training phase. It's made to "read" an astronomical amount of text: books, articles, web pages, conversations. No one dictates grammar rules or history lessons to it. Through sheer volume of examples, it spots patterns on its own: which words go together, how a sentence is built, how a line of reasoning unfolds, what a recipe or an official letter looks like.
It's a bit like a musician who has listened to thousands of pieces: they haven't memorised every note, but they've internalised a style and can now improvise. The AI, for its part, has internalised the regularities of language.
For images, the principle is similar: it's shown millions of images along with their descriptions, and it learns to link words to shapes, colours and textures.
An important point: it predicts, it doesn't "know"
This is probably the most useful idea to hold on to. Generative AI doesn't "understand" the world the way you and I do, and it doesn't consult a database of truths. At every moment, it calculates what is plausible.
Most of the time, what's plausible is also accurate — which is why its answers are often stunning. But not always. Sometimes it produces a perfectly worded statement… that's false. This is called a "hallucination". It might, for instance, invent a quote, a date or a book reference that doesn't exist, simply because it "sounded right".
This isn't a rare, embarrassing bug: it's a direct consequence of how it works. Hence a golden rule: generative AI is a wonderful assistant, not a source of authority. For anything that really matters (a figure, a piece of medical, legal or financial information), check it.
What is it actually good for?
Once you understand its limits, it becomes a surprisingly powerful tool for:
- Writing and rephrasing: emails, summaries, drafts, style corrections.
- Explaining: making a complicated topic accessible, translating, giving examples.
- Generating ideas: product names, article outlines, angles you wouldn't have thought of.
- Saving time on tedious work: sorting information, structuring messy notes, creating a first draft.
What all these uses have in common: the AI does the bulk of the initial work, and you keep control of the final decision.
In short
Generative AI is a machine that has learned the regularities of language (or of images) by observing huge amounts of examples, and uses them to produce new content, word after word, by looking for what is most likely. It's fast, creative and versatile, but it can be confidently wrong. The right attitude: treat it as a brilliant assistant that you proofread.
In the next article, we take a closer look at the engine that powers ChatGPT, Claude or LeChat: the LLM. What exactly is it, and why that odd-sounding name? We explain it all, simply.