It doesn't understand. It guesses.

The most useful thing to understand about large language models is also the most counterintuitive: they don't actually know anything.

What they do is predict the next likely word (technically a "token", but let's not go there) based on patterns learned from a truly enormous amount of text. That's the whole trick. You ask a question, the model figures out what words would plausibly follow given everything it has seen, and it produces them with the quiet confidence of someone who has read every book in every library, retained approximately none of it, but developed a very convincing way of sounding like they have.

This matters for a few reasons.

It explains hallucinations. When a model confidently tells you that the capital of Australia is Sydney, or invents a scientific paper complete with plausible-sounding authors and a journal that does not exist, it isn't lying. It's doing exactly what it was built to do: producing plausible-sounding text. The problem is that "plausible-sounding" and "accurate" overlap most of the time, but not always, and the model has no way to tell the difference.

It explains why the tone is always so certain. There's no hesitation baked into the prediction process the way there is in human reasoning. A person who doesn't know something will often hedge, trail off, or admit it. A language model just answers.

And it explains why treating it like a search engine is a mistake. A search engine finds text that exists somewhere. A language model produces text that should exist, based on what it has seen. Very different things.

None of this makes these tools useless. Quite the opposite, actually. But it does mean the mental model you bring to them matters a lot. Think of it less as asking an expert and more as having a very well-read assistant who sometimes makes things up and, crucially, cannot tell when they're doing it.

Useful, strange, and a little unsettling. Which, honestly, describes a lot of technology.