What the tool is, and what it isn't

You are probably already using an AI tool at work. Having a text rewritten, a long document summarised, an email put into English. It often works surprisingly well, and that is exactly why so little thought goes into it.

This course is not about everything you can do with it. It is about the four questions underneath: where do my words end up, is what I get back correct, whose is it, and what do we agree on here. Everything in this course traces back to what kind of thing such a tool actually is. So that is where we start.

It guesses the next word

A language model is trained on enormous quantities of text. What it has learned is which words tend to follow which others. When you ask a question, it does not look up an answer — word by word, it assembles the text most likely to follow your question.

That sounds like a technicality, but it explains almost everything else in this course. An answer that is correct and an answer that is invented are produced by exactly the same process. There is no step in which the model checks whether what it is writing is true.

Think of someone who has read an enormous amount and speaks very fluently, but looks nothing up and never says they don't know. You have probably met people like that. You use them to think out loud with, not as a source.

So: not a knowledge base, not a calculator, not an archive

Three things follow from that one property. It is not a knowledge base: it looks nothing up in a reference work, not even when it cites sources. It is not a calculator: to a language model, numbers are simply text, and a sum that looks right can be plain wrong. And it is not an archive of your organisation: what it says about your company is inferred from what has been written about companies like yours in general.

Some tools are now built more cleverly: they really do search the web, or work out a sum in a separate program. That helps, but it does not change the core. The text you end up reading is still predicted — including the sentence describing what the source supposedly said.

Why it always sounds certain

The hard part is not that an AI tool makes mistakes. The hard part is that you cannot tell from the answer whether it contains one. A colleague hedges: 'I think that's how it works, but do check.' A language model writes the most likely text, and the most likely text is a fluent, confident answer.

So you get no signal telling you when to pay attention. That means you have to build in a moment of checking yourself. Module 2 covers that, and that routine is the single most important habit in this course.

What this means for your work

The practical conclusion is simple: use an AI tool for work where you can judge whether the result is right. Rewriting a text, summarising a document you have read yourself, producing a first draft that you then adjust — there you are the expert and the tool is the instrument.

The moment you ask for something you cannot judge yourself — a date, an amount, a clause in an agreement, the name of a supplier — the roles change, and you have to verify it elsewhere. Not because it is bound to be wrong, but because you cannot tell from the answer whether it is right.