Hallucination occurs when a model generates text that sounds confident but is factually wrong or fabricated, often because it's pattern-matching rather than retrieving verified facts. Hallucinations can range from subtle inaccuracies to entirely invented citations, people, or events. They arise because language models optimize for plausibility, not truth. Mitigation strategies include retrieval-augmented generation, chain-of-thought prompting, and training with human feedback, but no current technique eliminates hallucinations entirely.