Fluent output that is simply not true.
A hallucination is a confident, well-formed statement with no basis in the model's inputs or in fact. It is a property of how these models generate, not a bug that can be patched out.
Models predict plausible continuations; plausibility and truth mostly overlap, and where they do not, the output is still plausible.
Mitigations are structural: retrieval so answers are grounded in real documents, citation so claims can be checked, and human review wherever the cost of a wrong answer is real.