What is a large language model?

A large language model is a program trained on a very large amount of text. It doesn’t look facts up. It predicts the next most likely piece of writing, one chunk at a time. That single idea explains most of what feels surprising about these tools — the fluency and the mistakes both.

When you type a question, the model isn’t searching a database for the answer. It is continuing your text in the way its training suggests is most probable. Most of the time the most probable continuation is also a useful, accurate one. Sometimes it isn’t.

Why it sometimes invents things

Because it predicts plausible text rather than retrieving verified records, it will sometimes produce something that reads well but isn’t true. People call this “hallucination”. It isn’t lying — there is no intent — it is the same machinery that makes the tool fluent, working on a question where the most fluent answer happens to be wrong.

Treat an assistant like a fast, confident junior colleague: useful, quick, and to be checked.

What it’s genuinely good at

Drafting, rephrasing, summarising, explaining a concept three different ways, translating rough notes into clean prose. These are tasks where “plausible, fluent text” is exactly what you want, and where you can read the result and judge it yourself.

It is weaker wherever being right matters more than sounding right: precise figures, recent events, legal or medical specifics, anything you can’t verify at a glance.

How to hold it in your head

You don’t need the maths. You need one habit: read what it gives you as a draft from a capable assistant who has not checked their sources. Keep the parts you can confirm, fix the rest. From there, everything else — better prompting, choosing a tool, using it safely — follows naturally.