Hallucination is when an LLM generates confident, fluent text that is factually incorrect — inventing citations, people, statistics, or events that do not exist. It is a fundamental limitation of how LLMs work: they generate plausible continuations of text, not verified truth.
Hallucination is the primary risk in high-stakes LLM deployments — legal, medical, financial. Mitigation strategies include RAG (grounding answers in retrieved documents), structured output validation, tool-calling to query live data, and human-in-the-loop review. Hallucination rates vary significantly by model, task type, and how well the prompt constrains the output space.