What is an AI hallucination?

In plain English

Language models generate text by predicting what plausibly comes next, not by looking facts up in a database. Most of the time plausible and true line up. When they do not, the model produces a confident, fluent, wrong answer: a hallucination. It might invent a book title, a legal precedent, a company policy or a web link that leads nowhere.

Crucially, nothing in the output signals the difference. Invented details are written in exactly the same assured tone as well-established ones.

Why it matters

Hallucinations are the single biggest practical risk in professional use of AI. People have filed legal documents citing cases that never existed and published articles with fabricated quotes. Because the errors are specific and plausible, they survive a quick skim and are caught only by someone who checks the source.

An example

A student asks for five academic references on a niche topic. Three are real, one has the right authors but the wrong journal, and one is a complete invention with a convincing title, volume number and page range. The list looks uniform and authoritative; only searching each entry reveals the problem.

What to watch out for

Asking a model whether it is sure does not help much — it can be just as confident when wrong, and it may apologise and invent something new. Features that let the model search the web or read your documents reduce hallucination but do not eliminate it; it can still misread or over-generalise a source. The practical defence is unchanged: treat every name, number, quote and citation as unverified until you have checked it yourself.