Hallucination
A confidently-stated but false claim produced by an AI model — a fabricated citation, an invented statistic, a made-up event.
A hallucination is what happens when a language model outputs something that reads like a fact but isn't. Wrong dates, invented papers, non-existent people, misattributed quotes. The word is metaphorical — nothing is actually being "seen" — but it captures that the model isn't lying (which implies intent to deceive) and isn't merely mistaken (which implies checking against a source of truth). It's producing statistically plausible text that happens to be false.
Hallucinations happen because language models are trained to produce probable continuations, not verified ones. When asked something they don't reliably know, they don't say "I don't know" — they interpolate, and the interpolation is often wrong in specific ways that look confident.
In practice: for any claim that matters (a citation you'll publish, a date you'll act on, a name you'll email), verify. Use cited web search where available. Ask the model to flag its own confidence. Treat unfamiliar specifics as suspects, not sources.