What AI is, where you meet it, and why it gets things wrong

Most software follows instructions a person wrote. A calculator adds two numbers because somebody wrote down exactly how to add. It does the same thing every time, and if it is wrong, someone wrote the instruction wrongly.

AI works differently. Nobody writes down the rule. Instead the machine is shown an enormous number of examples and works out the patterns for itself. Show it a million photos labelled 'cat' and it learns what tends to appear in a cat photo.

That is the whole trick, and it explains almost everything else about AI. It is very good at things that have patterns, and unreliable at things that need actual understanding.

It also means AI does not know anything in the way you know things. It has noticed what usually goes together. That is not the same as understanding, and the difference matters when it is wrong.

Ordinary software follows rules someone wrote. AI finds patterns in examples.

When a video app decides what to show you next, that is AI. When your photos app finds every picture of your dog, that is AI. When a bank blocks a payment as suspicious, when a map reroutes you around traffic, when your keyboard guesses the next word — all AI.

Chatbots are the newest and loudest kind, so people assume they are the only kind. They are a small part of it.

This matters because you cannot decide how you feel about AI if you think you have never used it. You almost certainly have, dozens of times today.

It also means AI arrives without being announced. A tool you have used for years can gain an AI feature in an update, with nothing more than a new icon.

AI is already in recommendations, photos, maps, keyboards and fraud checks — not just chatbots.

An AI system is trained by being shown examples. To learn to spot spam email it is shown a great many emails, each labelled spam or not spam, until it picks up what spam tends to look like.

So the examples decide what it learns. Show it only photographs of one kind of dog and it will struggle with every other kind. Show it examples full of mistakes and it will learn the mistakes.

This is why people say 'rubbish in, rubbish out'. The machine has no way of knowing that its examples were poor. It simply learns them.

And once training is finished, the system does not keep learning from you as you use it. It applies what it already learned. That is why a tool can repeat the same mistake even after you tell it it is wrong.

AI learns from the examples it is given, including their mistakes and gaps.

A chatbot will sometimes state something completely untrue in the same calm, fluent voice it uses for everything else. It may invent a fact, a date, a book that was never written, or a person who does not exist.

This is not the system lying, because lying needs an intention. It is producing text that fits the pattern of a good answer. Text that reads like a correct answer and text that is a correct answer are not the same thing, and the machine is aiming at the first one.

So confidence tells you nothing. A wrong answer sounds exactly as sure as a right one. There is no wobble in its voice to warn you.

The habit that protects you is simple: if it matters, check it somewhere else. Not because AI is useless, but because you cannot tell which answers are the wrong ones by reading them.

AI can be confidently wrong, and it sounds the same either way. If it matters, check it.

When you type something into an AI tool, it usually travels to a company's computers. Depending on the tool and its settings, it may be stored, reviewed by staff, or used to improve future versions.

So the sensible rule is the one you would use for a postcard. Do not write anything you would mind a stranger reading.

That means no passwords, no bank details, no home address, no medical details, and nothing private about somebody else — your friend's problem is not yours to hand over.

Work and school matter too. Pasting a private document into a public AI tool can breach a rule you did not know existed. If in doubt, ask before pasting, not afterwards.

Treat anything you type into AI as a postcard a stranger might read.

AI can help you understand something difficult, practise a language, plan a piece of work, or get past a blank page. Those are good uses and nobody sensible objects to them.

The trouble starts when AI writes something and you present it as your own. In school that is usually cheating. At work it can be a sacking. Either way the problem is the pretending, not the tool.

The straightforward fix is to say what you used. 'I used AI to draft this and then checked and rewrote it' is an honest sentence, and most rules allow it.

One more thing. AI can now make convincing fake photos, voices and videos of real people. Making one of somebody without asking is a way of hurting them, and in many places it is a crime. If you would not do it to their face, do not do it with a tool.

Say when you used AI. Do not make fake images or voices of real people.