Hallucination
When a model states something fluent and confident that is simply not true.
A hallucination is a plausible-sounding but false or unsupported output. Because models predict likely text rather than look up facts, they can invent citations, APIs, or numbers. Grounding (RAG), tool use, and evals are the main mitigations.
Worked example: ask for a source and a model may produce a perfectly formatted citation — real-sounding title, plausible authors, a journal, a year — for a paper that does not exist; ask about a library and it invents a method that should exist by analogy but does not. Gotcha: fluency is uncorrelated with truth — the model sounds exactly as confident when it is wrong as when it is right, so you cannot detect a hallucination from tone. That is why the fix is structural (retrieve real sources, call real tools, verify with evals), not “tell it to be accurate”.