Hallucination

A hallucination is any part of an AI's answer that it was not given and is not true: an invented statistic, a made-up policy detail, a citation to a document that does not exist. The wording usually reads just as confidently as the correct parts, which is what makes it dangerous.

Grounding the model in retrieved sources reduces hallucination but does not erase it, so the reliable defense is measurement: check each claim in an answer against the source it came from. In information terms, a hallucination is the model filling in uncertainty the question left unspecified.