When an AI confidently says something false or made up as if it were true.
A hallucination is when an AI states false facts, wrong numbers, or books and papers that don't exist, as if they were real. It isn't lying on purpose. It picks likely-sounding next tokens instead of looking facts up, so it can string together something smooth but wrong.
It happens most with rare topics, recent news, exact numbers and sources. A confident tone is not evidence.
RAG (giving the AI real documents to read first) and tool use (letting it search or calculate) help, but nothing removes it completely. For anything important, check another source yourself.
Think of it like this
Picture a student who didn't study but won't leave an essay question blank. They fill the space with confident, nice-sounding sentences, and some of it is simply wrong. Neat handwriting doesn't make an answer correct.
For example
Ask for "a study that proves this" and an AI may invent a realistic title and author names.