What can't AI do? Limits and how to check the facts

Why it pays to know the limits

AI writes fluently, quickly and in a confident voice. That's exactly what makes its mistakes treacherous: wrong information looks exactly like right information. Nobody would sign a document without reading it. The same goes for text from an AI.

This guide isn't a warning against using AI. It's a user manual: once you know the limits, you know when you can trust an answer and when you need to check it.

1. Hallucinations: AI makes things up

A hallucination is when an AI invents information that doesn't exist and presents it as true. It might invent:

• a law or a government benefit that doesn't exist
• a book, study or newspaper article that was never published
• a company's phone number, address or opening hours
• a quotation nobody ever said

Why does this happen? A language model is trained to produce believable text. When it doesn't remember something, it fills the gap with something that sounds right. It isn't lying on purpose, but the result is still wrong. More background in how AI works.

Newer models make fewer mistakes, and many can now say "I'm not sure". But the problem hasn't gone away, so checking still matters.

2. Outdated information

An AI model is trained up to a certain date. It doesn't automatically know what has happened since: new laws, changed prices, election results or whether your local shop has changed its opening hours.

This matters especially for things like tax rates, benefits, energy prices and pensions, whose amounts and rules often change at the start of a new year. An AI might give you last year's figure as if it applied today.

Some services can search the web and will usually show links to their sources. Perplexity, for example, is built specifically to search and show sources. Even then, it's worth opening the link and checking that the source really says the same thing.

3. Maths and exact figures

It sounds surprising, but a computer program can get sums wrong. The reason is that a language model treats numbers as text rather than calculating like a calculator. Simple sums usually come out right, but long calculations, percentages, interest and date differences can go wrong.

Many services now use a separate calculation tool behind the scenes, which has improved things noticeably. Still, if you're working out a loan repayment, a VAT amount or a budget, check the result with a calculator or spreadsheet. More in Excel and data.

4. Sources and references

If you ask an AI for its sources, it may give you a convincing-looking list – but some links may not work, or the articles may not exist. This applies especially to services that don't search the web.

Rule of thumb: a source only counts as a source once you've opened it yourself and seen that it says what the AI claims.

5. Things AI simply can't know

AI knows nothing about your own situation unless you tell it. It can't see your medical records, your tax code, your bank account or your business accounts. It doesn't know what your doctor told you last week. So general advice won't always fit your particular case.

Nor can it judge things that need your senses and a visit in person: whether the cottage roof really is in poor shape, whether the cellar smells of damp, or whether a used car is a good buy.

How to check an AI answer – five steps

1. Ask how sure it is. "How confident are you about this? Which parts should I double-check?" AI can often point out the weak spots itself.
2. Ask for sources and open them. Don't settle for a list of links – open at least the most important ones.
3. Check the official website. Your tax authority, social security office, health service, local council or the company's own site is always more reliable than an AI answering from memory.
4. Rephrase the question or ask another AI. If the answers contradict each other, at least one is wrong.
5. Use common sense. If something sounds too good or too strange to be true, it often is. Life experience is the best tool here.

Try it: ask the AI to check itself

When you've received an important answer, paste this follow-up into the same conversation:

This doesn't automatically make the answer correct, but it highlights the points you should verify.

Where AI is good – and where it isn't

Summary

AI can make things up, give outdated information, get sums wrong and offer sources that don't exist. It also knows nothing about your situation unless you tell it. That doesn't make it useless – it makes it a tool to be used with thought. Use AI for drafting, explaining and brainstorming, and always check figures, laws and health matters with official sources. For tips on asking better questions, see how to ask AI well.