How does AI "know" things? Where the answers come from

Ask a chatbot about Roman history, sourdough bread or the rules of cricket, and it answers instantly, in full sentences, as if it had read everything ever written. So the natural question is: how? Is it looking things up? Is someone typing on the other end? Does it have the internet inside it?

None of those. The real answer is stranger, and once you understand it, a lot of AI’s behaviour — the brilliance and the mistakes — makes sense.

It learned by reading. An enormous amount.

Before you ever typed a word to it, the model (the engine behind the chatbot) was “trained.” That means it was shown a staggering quantity of text — most of the public internet, millions of books, articles, forum posts, recipes, manuals — and given one job, over and over, billions of times:

Guess the next word.

Show it “The capital of France is…” and it guesses “Paris.” Get it wrong, adjust, try again. Repeat for months on thousands of powerful computers. By the end, it has become extraordinarily good at predicting what words are likely to follow other words — about almost any subject humans have written about.

That’s it. That’s the whole trick. Everything it does grows out of that one skill.

Why that makes it feel like it “knows”

If you’ve read enough about cricket, you can explain cricket. Not because you memorised a rulebook, but because the patterns are in you. The model is the same, at a scale no person could match. When you ask about cricket, it isn’t fetching a stored answer — it’s composing one, word by word, from the patterns it absorbed.

That’s why it can explain the same thing five different ways, adjust for a ten-year-old or a professor, or write it as a poem. It isn’t retrieving; it’s generating.

What it doesn’t have

This is the part that matters for using it well.

It doesn’t have a library card. Unless the chatbot has been given a search tool (some have — they’ll show you links), it cannot check a fact. It can only produce what sounds right based on patterns. Usually that’s also what is right. Not always. When the patterns are thin — a small town, an obscure book, a recent event — it will still confidently generate something. That’s a hallucination, and it’s the single most important thing to know about AI.

It doesn’t know what happened last week. Training has a cut-off date. Anything after that, the model has never seen. Ask about this morning’s news and it will either tell you it doesn’t know, or — worse — guess. Chatbots with search can get around this; ones without can’t.

It doesn’t know you. It cannot see your screen, your other tabs, your files or your previous chats (unless the product has a memory feature, and it will say so). Unless a memory feature is switched on, every new conversation starts cold (more on that here). If you want it to help with your letter, you have to paste the letter in.

It doesn’t have opinions or intentions. When it says “I think…”, that’s a figure of speech learned from human writing. There’s no one in there thinking. This is comforting when it’s rude and important when it’s flattering: neither means anything.

“But it got my question exactly right”

It often will. Common questions have been asked and answered thousands of times in the training text, so the patterns are strong and the answer is reliable. “How do I boil an egg?” is safe. “What are the side effects of my specific medication at my dose?” is not — the patterns are thinner, the stakes are higher, and it will sound equally confident either way.

A useful rule of thumb: the more people have written about it, the more you can trust the answer. The more specific, personal or recent your question, the more you should check.

So where should you use it?

Once you know it’s a pattern-machine, not an oracle, the good uses become obvious:

• Explaining things you could check elsewhere: “explain compound interest like I’m twelve.”
• Drafting anything you’ll read before sending: emails, letters, speeches. (See writing a difficult email.)
• Rewording what you paste in: “make this shorter,” “make this more polite,” “what does this letter actually say?”
• Brainstorming: gift ideas, questions to ask a doctor, names for a cat.

And the uses to be careful with: anything where a confident wrong answer would cost you — money, health, legal matters, or a fact you’re about to repeat to someone else. Use it to prepare for those; don’t let it have the last word.

The one-sentence version

AI answers by predicting words, based on patterns from an enormous amount of human writing — which is why it’s fluent about everything, right about most things, and unable to tell you which is which.