What is an AI hallucination and how do you spot it?
A hallucination is an invented fact presented with exactly the confidence of a true one: a quote nobody said, a clause of law that does not exist, a round number that sounds about right. You catch it by asking for the exact sentence that backs each claim and by checking everything specific by hand — figures, dates, proper names and references.
Why it happens
Because the model is aiming for a continuation that fits, not one that is true. When you ask for a fact it does not hold, the most likely continuation is not "I do not know": it is something shaped like a fact. A plausible number, a plausible date, a plausible name. That is why inventions look so good.
Where it invents most
- Figures and statistics, especially round ones or anything "according to a study".
- Quotes and references: books, rulings, clauses, links.
- Recent events, later than whatever the tool knows about.
- Very local detail: opening hours, prices, the paperwork in your town.
- Anything you push for twice when it plainly does not have it.
Three defences that cost nothing
- Supply the material yourself. Paste the document and say to work only from that, and there is barely room to invent.
- Ask for the exact sentence. "For each point, copy the sentence from the text that supports it." No sentence, no fact.
- Authorise "I do not know". "If it is not in the text, write: not stated." It sounds obvious and it changes the output.
The two-minute check
Before you forward anything, underline the specifics — figures, dates, names, references — and check only those. That is where about ninety per cent of the risk sits, and it takes a couple of minutes. The rest of the text, if the argument holds, you can read with ordinary judgement.
One warning sign covers the lot: the more confident an answer sounds about something very specific that you did not supply, the more it needs checking.
An AI does not lie; it fills in. Your job is not to sign off what it filled in without looking.