A hallucination is output that is false or unsupported by the provided context, stated confidently. Hallucinations are intrinsic to how models generate likely text, so the goal is detection and reduction, not elimination: ground answers in retrieved context, prompt for citations, and score faithfulness with evaluators.
Why it matters
Hallucinations are fluent and confident, so they are hard for users to spot and can do real damage when acted on. They are intrinsic to how models generate text, not a bug you fix once.
How it works
The model predicts likely text, not verified truth, so it can invent facts or citations. The practical defenses are grounding answers in retrieved context, prompting for citations, and scoring faithfulness with evaluators.