hallucination
/ huh-loo-suh-NAY-shun /
A hallucination is when a language model states something false, made-up, or unsupported with the same calm confidence it uses for true things. It invents a plausible-sounding court case that never happened, cites a study with a real-looking author and a fake title, or confidently misremembers a date. Crucially, it does not signal any doubt — the made-up fact arrives in the same fluent, authoritative voice as a correct one, which is exactly what makes hallucinations dangerous.
The reason this happens is baked into how the models work. A language model is trained to produce text that is statistically plausible given everything before it — not text that is verified against reality. It has no separate store of facts to check against and no built-in sense of "I don't actually know this." When it lacks the real answer, the most plausible continuation is often a confident, well-formed guess. The model is not lying — lying requires knowing the truth and choosing to hide it; the model is simply generating likely words, and sometimes the likely words are wrong.
This is not a bug that a future update will quietly fix; it is a fundamental consequence of training models to predict plausible text. Techniques help — retrieval-augmented generation grounds answers in real documents, and asking for sources or expressing uncertainty can reduce it — but nothing eliminates it. The practical upshot is unglamorous but essential: for anything that matters, treat a model's confident factual claim as a lead to verify, not as an answer to trust.
Asked for sources on a niche topic, a model produces three references — correct journal names, realistic author lists, plausible titles, even page numbers. Two of the three papers do not exist. The format is impeccable; the content is invented. Several lawyers have been sanctioned for filing exactly such fabricated citations.
Perfect format, invented content — the hallmark of a hallucination.
"Hallucination" is partly a misleading word: the model isn't malfunctioning, it is doing exactly what it was built to do — generate plausible text. The mismatch is in our expectation that plausible equals true. Hallucination cannot be fully removed; it can only be reduced and caught.