Hallucination
A fluent, confident, entirely invented answer. The failure mode that matters, because it looks exactly like a correct one.
Also written: hallucinate, hallucinations, confabulation.
A model that does not know something does not usually say so. It produces the most plausible continuation, and a plausible-looking citation, function name or statistic is often one that does not exist.
This is not a bug awaiting a patch; it follows from what the thing is. A next-token predictor has no separate store of facts to check itself against, and no internal signal that distinguishes recall from invention.
What works is structural, not verbal: give the model the source material rather than relying on its memory, ask for citations you can click, and check anything load-bearing. Telling a model not to hallucinate does approximately nothing.