Hallucination is when a large language model produces confident, well-formed output that is factually wrong. Fake citations, invented case law, wrong dates, non-existent people, and made-up API endpoints are all hallucinations.
Why it happens: LLMs are trained to predict the most likely next token, not to verify truth. When they lack information, they default to plausible-sounding completions rather than admitting uncertainty.
Mitigations: retrieval-augmented generation (grounding answers in real documents), structured output constraints, self-consistency checks, chain-of-thought reasoning, and human-in-the-loop review for high-stakes decisions. In legal, medical, and financial applications, we consider hallucination unmitigated risk unless there is a verification layer.