Hallucination
An answer that is fluent, confident, and wrong — a fact, a citation, or an API that does not exist.
The model predicts what a plausible answer looks like, not what is true. Where the training data is thin — a version number, a person's title, a rarely used function — the plausible shape gets filled with invented content, and it reads exactly like the rest.
Nothing removes it entirely. What works is narrowing the room for it: give the model the source instead of asking it to recall, ask for the quote alongside the claim, and check anything with a number in it against the primary source before it is published. That last rule is why every number on this site carries a date.