What is an AI Hallucination?

An AI hallucination is when an AI model states false information with complete confidence: invented facts, fake citations, events that never happened. It happens because the model generates text that sounds plausible rather than looking facts up. Wrong answers read exactly like right ones.

Hallucinations happen because a language model predicts the most plausible next words based on patterns in its training data, and plausible is not the same as true. When the model knows a topic well, what sounds right usually is right. When it does not, the words keep flowing anyway, just as fluent and confident, now filled with things that merely sound right: a fake court case, an invented statistic, a book the author never wrote. The famous early example was lawyers sanctioned for filing a brief citing six court cases the AI simply made up. Two things make it dangerous: the confidence never drops, so you cannot hear the difference, and the errors land hardest exactly where checking is tedious, in citations, numbers, and technical details. Fixes reduce it rather than cure it: grounding the model in retrieved documents so it quotes instead of recalls, asking for sources, lowering how creatively it samples, and keeping a human check on anything that matters.

The brilliant surgeon at the family dinner. On medicine she is genuinely superb, and that is exactly the problem. The same authoritative voice covers car engines, tax law, and ancient Rome, topics she knows a tenth as well, because decades of being right have erased her sense of where her expertise ends; she is not lying, she cannot feel the boundary. Everything past it comes out in the same confident surgeon voice, and the guests only learn which claims were solid when they check later.