What is a model, really?

An AI model — the kind behind ChatGPT, Claude, Gemini and the rest — is a very large statistical pattern-matcher. That is not a simplification for beginners. It is what it is.

The one-sentence version

It has read an enormous amount of text, and it produces the continuation that best matches the patterns it learned.

When you type a question, the model is not looking anything up. It is generating a response one small piece at a time — each piece chosen because it is what tends to come next, given everything so far.

Why that explains almost everything

Once you know the mechanism, most strange AI behaviour stops being mysterious:

It sounds confident because confidence is what the training data sounds like. Text written by people who knew what they were talking about does not hedge much. The model learned the tone without the knowledge.

It invents things because it is built to produce something. It has no mechanism for saying "I do not know" — so when it lacks the pattern, it produces the most plausible-sounding continuation instead. This is called hallucination, and it is a feature of the design, not a bug being fixed.

It cannot know about your company, your files, or last week. It learned from text collected up to some point, and it has no memory of your conversation beyond the current one.

It is not searching the internet unless something has explicitly been added to make it search.

The analogy that mostly works

Think of the autocomplete on a phone keyboard, which suggests the next word, except trained on a library the size of a small country — and then scaled up until the suggestions became useful for whole answers rather than single words.

That is genuinely close to what is happening. It also shows why the results feel so human and fail so strangely: autocomplete is very good at sounding like English, and that is not the same as being right.

What that means for you, practically

The model is a language engine, not a knowledge engine. That has one immediate consequence, and it is the most important habit in this whole section:

Treat its output as a fluent first draft, never as a source.

Fluent and correct are different properties, and the model is much better at the first one. Everything else — the prompting, the checking, the tools — exists to close that gap.

Where to go next

• What it can and cannot do — honest capabilities, without hype in either direction
• The words people use — the vocabulary, explained plainly
• How to check what it tells you — the habit that matters most