What AI actually does, in plain words

The single most useful thing you can know about tools like ChatGPT, Claude and Gemini is what kind of machine they are. Not how they were built. What they are doing when you ask them something.

Here it is: they predict what text should come next, one small piece at a time, based on patterns in an enormous amount of writing they were trained on.

That sounds reductive, and it explains almost everything about how they behave.

Why that one fact explains so much

It explains why they are brilliant at shaping language. Rewriting, summarising, changing tone, turning notes into an email, turning an email into notes. This is the thing they do natively, and they are better at it than most people are.

It explains why they invent facts. A model producing plausible next words does not know the difference between a real conference and a convincing-sounding one. When it does not have the answer, it does not stop. It produces the shape of an answer. This is usually called hallucination, and it is not a bug being fixed next year. It is a property of the machine.

It explains why they are confident when wrong. Confidence is a style of writing. The model reproduces that style whether or not the content underneath is right.

And it explains why the request matters so much. You are not searching a database. You are setting the conditions for what gets produced. Vague conditions, generic output.

What they are genuinely reliable for

I use these tools every working day. The jobs they never let me down on:

Turning something long into something short. Meeting notes, a thread, a report you were sent and do not have an hour for.
Turning something short into something long. Bullet points into a first draft, an outline into a section.
Changing register. The same message written for a customer, for your manager, and for a colleague you are friendly with.
Being a first reader. "What is unclear here" and "what would a sceptical reader push back on" both work extremely well.
Structuring a mess. Fifteen unordered thoughts into four themes.
Explaining something at your level. Ask it to explain a contract clause as if you have never read one, and it will.

Notice the pattern: every one of those is a language job where you can check the output yourself. That is the safe zone.

Where they fail, reliably

Anything requiring current facts. Unless the tool is actively searching the web, its knowledge stops at a cut-off date and it will not warn you.
Numbers. Arithmetic across a table, percentages, dates and durations. Check every figure. I have been burned by a confidently wrong total more than once.
Anything about you it was not told. It does not know your company, your customer or last week's decision unless you paste it in.
Citations. Asking for sources without web search on produces sources shaped like real ones. Some will not exist.
Judgement calls with consequences. Legal, medical, financial. It will answer. That is exactly the problem.

The mental model that has held up

Treat it as a fast, widely-read colleague who has never met your company, will never say "I don't know", and needs everything checked. You would still hand that person plenty of work. You would not send their output to a customer unread.

That framing gets you most of the value with none of the disasters.

Two words worth knowing, and no more

Model is the thing doing the predicting. GPT-5, Claude, Gemini: those are models. Newer usually means better at reasoning and more expensive to run.

Context is everything it can see right now: your message, the conversation so far, and any file you attached. It has no memory beyond that unless the product adds one. This is why a fresh chat forgets last week, and why pasting the actual document beats describing it.

That is the whole vocabulary you need. Everything else is marketing.

What to do next

Open whichever tool you already have access to and give it a real piece of work you finished last week. Ask it to critique it. You will learn more from that in five minutes than from another ten explainers.

Then read how to ask an AI for something useful, because the request is where nearly all the quality comes from.