Hallucination

A hallucination is output from a language model that is fluent, plausible and wrong — an invented citation, a function that does not exist, a confidently stated figure with no basis. It is not a bug in a particular model but a consequence of how these models work: they produce likely continuations, and a likely-sounding falsehood scores well.

This is why the deployment question is never "will it hallucinate" but "what happens when it does". A drafting tool where a person reviews every output has a low cost of error. A system that files something, emails a customer or acts on its own conclusion has a high one, and needs grounding, citations and a verification step to match.

Retrieval-augmented generation reduces hallucination substantially by putting the real source text in front of the model. It does not eliminate it, and any interface that shows an answer without a way to check it against the source is inviting the failure it is trying to avoid.