Hallucination is when a language model generates output that sounds authoritative but is factually wrong or fabricated. The model has no awareness that it's wrong — it's not lying, it's just confidently mistaken.

This is a structural property of how LLMs work, not a bug that will be patched away. These models predict statistically plausible next tokens. Plausible and true overlap a lot, but not always. In the gap between those two things live invented citations, wrong dates, nonexistent APIs, and confident nonsense about your specific domain.

Mitigation is an engineering discipline, not a setting you toggle. RAG grounds model answers in real retrieved data rather than parameterized memory. Evals quantify your actual error rates so you know what you're shipping. Guardrails catch bad outputs before users see them. None of these eliminates hallucination — they bound and monitor it.

Teams that treat hallucination as a dealbreaker never ship. Teams that treat it as no big deal ship systems that erode user trust. The professional position is in the middle: measure it, bound it operationally, and design the product so the cost of an error is manageable when it happens. Because it will.