How it works, what it does to what you see, and what you can ask for

There are two separate stages. Training is when the system is built: it processes an enormous body of data and adjusts millions of internal settings until its outputs match the examples. This takes months and costs a fortune. Using it afterwards is fast and cheap.

A large language model — the thing behind a chatbot — works by predicting what text should come next. Given everything so far, it estimates the most likely continuation, adds it, and repeats. That is genuinely all it is doing.

This explains the strangest thing about them. Fluency and accuracy are separate properties. A model can produce beautifully written text that is entirely false, because 'reads well' and 'is true' are different targets and it is aiming at the first.

It also explains why the same question can produce different answers. There is an element of randomness in how the next word is chosen, which is what stops the output being flat and repetitive.

A language model predicts likely text. Fluent and true are different things.

A recommender learns from what people actually do — what they watch to the end, replay, react to — and predicts what will hold your attention next. It is measured on engagement, because engagement is what can be counted.

The problem is that what holds attention is not the same as what is true, useful, or good for you. Outrage holds attention well. So does content that confirms what you already think.

The result is a feed shaped around a version of you built from your behaviour, not your intentions. Two people searching the same term can see genuinely different worlds, and neither is told that.

You are not powerless. What you watch to the end, search for and engage with is the input. Deliberately seeking out sources that disagree with you changes what the system predicts you want.

Feeds optimise for attention, not accuracy. Your behaviour is the input.

When a model produces a confident, specific, completely false statement, that is usually called a hallucination. The word is poor — it suggests a malfunction — but the behaviour follows directly from predicting plausible text.

It is most likely where specifics are involved: numbers, dates, names, citations, quotations, legal provisions. These are exactly the details people copy without checking, because a precise-looking detail feels checked already.

Asking the model whether it is sure does not help. That question just generates more text, and the answer will sound equally confident either way. Some tools now cite sources, which helps only if you open them — invented citations are a known failure.

The working habit: use AI to draft, explain and explore, then verify any specific claim against a source that exists independently of the model. The more precise the claim, the more it needs checking.

Verify specifics — numbers, names, dates, citations — against an independent source.

Convincing fake images, voice and video can now be made cheaply by anyone. The advice to look for odd hands or strange blinking is obsolete — those artefacts were a temporary weakness, not a permanent tell.

Since you cannot reliably judge by looking, the better questions are about where it came from. Who published it? Does any independent source have the same footage? Does it arrive exactly when it would be most damaging?

There is a second effect, sometimes worse than the fakes. Once everyone knows video can be faked, anyone caught on camera can claim a real recording is a deepfake. The doubt becomes the weapon.

The law is moving. The EU requires AI-generated content to be disclosed and marked, several countries criminalise sexual deepfakes of real people, and rules on election material are tightening. The detail differs by country, but making non-consensual fakes of real people is increasingly illegal as well as wrong.

You cannot spot fakes by eye. Ask where it came from and whether anyone consented.

AI is genuinely useful for drafting, summarising, explaining, translating, practising and getting unstuck. Refusing to use it on principle is not a strategy.

Three habits keep that use legitimate. Say when you used it. Check anything it asserts as fact. And keep confidential material out of tools that were not approved for it.

The third catches people out most. Pasting a client document, unreleased results or someone's personal data into a public tool can breach a confidentiality duty or a data protection law, regardless of how careful you were afterwards.

Remember also that the responsibility stays with you. If you submit it, you own it. 'The AI produced it' has not worked as a defence for anyone who has tried it, in school or in court.

Disclose that you used it, verify what it claims, and keep confidential material out.

If a system learns from past decisions, it learns the patterns in those decisions — including the unfair ones. A hiring model trained on who was hired before will reproduce who was hired before. Nobody has to intend this for it to happen.

Removing the obvious field rarely fixes it. Delete gender and the system may still pick it up from a school, a sport, a gap in employment. The pattern survives the deletion of the label.

This matters because AI is used in decisions that affect people: job applications, credit, insurance, benefits, school places. These are not hypothetical uses.

You have rights here, particularly in Europe and Switzerland. Broadly: you can ask whether a decision about you was automated, ask for an explanation of how it was reached, ask for a human to look at it again, and ask what personal data an organisation holds about you. How to exercise them varies by country, but the right to ask is real and it is free.

Bias comes from the data. You can ask for a human to review a decision about you.