Hallucination

Hallucinations are LLMs' biggest production risk. The model generates plausible-sounding text that simply is not true — citing papers that do not exist, attributing quotes to wrong people, inventing API endpoints. The output sounds confident; the facts are made up.

Hallucinations happen because LLMs predict likely text, not retrieve facts. When the model has not seen reliable information about a topic, it fills the gap with plausible-sounding patterns rather than saying "I do not know".

Mitigation: (1) use **RAG** to ground answers in retrieved documents, (2) ask the model to cite its sources, (3) use a **smaller, more conservative model** for factual tasks, (4) post-process outputs against known facts, (5) tell the model it is OK to say "I do not know".