AI Hallucination

An AI hallucination is an answer generated by an artificial intelligence system that contains information that is false, unsupported or invented — even though the response may sound completely convincing.

The word hallucination is deliberately vivid. The system is producing something that is not there: a fact that doesn't check out, a citation that doesn't exist, a name or date that's plausible but wrong, a quote that was never said. Because the language is fluent and confident, it reads as if it came from a reliable source. It did not.

Hallucinations happen because of how models like the large language model behind ChatGPT actually work: they generate likely language one piece at a time, based on patterns learned from training. They are optimised to produce plausible-sounding text, not to verify whether that text is true. When the model doesn't know something, it may still produce a confident-sounding answer rather than admit uncertainty — especially if the prompt doesn't explicitly invite it to say so.

The forms this takes are varied. A model might invent a real-sounding book title and author, fabricate a legal citation, give a confident but wrong answer to a maths question, quote a study that doesn't exist, or describe a feature in a product that the product doesn't have. None of this is malicious; it is the model doing what it does — completing language in a likely way — in a situation where the likely completion happens to be wrong.

Confident language is not proof of accuracy. This is the single most important habit to build around AI: treat a fluent answer as a candidate, not a conclusion. For anything that matters — a figure you'll quote, a fact you'll act on, a reference you'll cite — check it against an authoritative source. Asking the model for its sources can help start that process, but the model's own claim of a source is itself something to verify, since it can invent those too.

Some situations carry higher risk and deserve extra care: anything legal, medical or financial; specific numbers, dates and names; quotations and citations; and instructions with real-world consequences. In those cases, independent verification isn't optional. Better prompting can reduce hallucinations — for example, by giving the model accurate context and inviting it to say when it is unsure — but it cannot eliminate them entirely. The model has no reliable internal sense of what it truly knows versus what it is generating plausibly, so the responsibility for catching errors stays with the person reading the answer.