Most explanations of AI start with neural networks and end with you no closer to using the thing. So let's skip that.

Here is the only mental model you actually need: these tools are extremely good at predicting what words should come next. That's it. That's the trick.

They were shown an enormous amount of writing, books, websites, manuals, arguments, recipes, legal documents, everything, and they got very, very good at working out what a sensible continuation looks like. When you type a question, the tool isn't looking up an answer in a database. It's producing what a well-informed response to your question would probably look like, one piece at a time.

Why that one fact explains almost everything

Once you understand that, the behaviour that confuses people starts making sense.

It explains why it's brilliant at some things

Rewriting an email to sound firmer. Summarising six pages into one. Turning your rambling voice note into a structured letter. Explaining a term you're embarrassed to admit you don't know. All of these are essentially "produce sensible words based on these words", which is exactly the thing it does.

It explains why it sometimes makes things up

If you ask for a case reference, a statistic, or a section of the Companies Act, it will produce something that looks exactly right, correct format, plausible numbers, confident tone, because a correct-looking answer is what usually follows that question. Whether it's true is a separate question, and one the tool is not really equipped to answer.

This is the single most important thing to carry with you. It is not lying, and it is not broken. It's doing its job, and its job is not the same as knowing.

It explains why how you ask matters so much

A vague question has a huge range of sensible continuations, so you get a vague, generic answer. A specific question with context narrows the field enormously. That's why the same tool gives one person mediocre results and another person excellent ones.

The one-line version

It's a very well-read assistant with an excellent turn of phrase, no memory of where it learned anything, and no ability to tell you when it's guessing.

What "generative AI" and "LLM" mean

You'll see these two terms constantly. They're less intimidating than they sound.

Generative AI means AI that produces something new (text, an image, a voice, a video) as opposed to AI that sorts or scores things. Your bank has used AI for fraud detection for years; that's not generative. ChatGPT writing you a letter is.

LLM stands for Large Language Model. It's the technical name for the kind of AI we've been describing. When someone says "which LLM do you use," they're asking which of ChatGPT, Claude, Gemini and the rest you prefer.

That's genuinely all the vocabulary you need to hold your own in a conversation about this.

What it isn't

  • It isn't a search engine. Some of these tools can now search the web, but the underlying thing is not looking up pages. If you need a source, ask for one and then go and check it.
  • It isn't a calculator. It's got better at arithmetic, but it's fundamentally producing likely-looking text. Check any number that matters.
  • It isn't conscious, and it isn't plotting. There's a lot of noise about this. For your purposes, it's a tool that produces text.
  • It isn't the same thing every time. Ask the identical question twice and you'll get two different answers. Both may be fine. This surprises people who expect software to be deterministic.

So what do you actually do with it?

The most reliable use is anything where you already know what good looks like. You can judge whether an email hits the right tone. You can tell whether a summary of your own document is accurate. You know if the tone of that client response is too soft.

The risky use is anything where you can't tell whether the answer is right. That's where it hands you a confident, well-written wrong answer and you have no way to catch it.

Use it where you're the expert.Be careful where you're not.

That single rule will keep you out of almost all the trouble people get into with these tools, and it's why the roles that get the most out of AI are experienced ones. Twenty years of knowing what a good client letter looks like isn't made worthless by this technology. It's the thing that makes you able to use it safely.