AI Hallucination

AI hallucination is the generation of plausible-sounding but factually incorrect, fabricated, or unverifiable content by a language model. Hallucinations occur because LLMs predict statistically likely token sequences rather than retrieving verified facts — they have no internal 'truth checker.'

Hallucination rates vary by model, task type, and prompt design. Mitigation strategies — in order of reliability — are: grounding to retrieved sources (RAG), structured output with citation requirements, constrained generation using schemas, chain-of-thought reasoning, and self-consistency checks. For high-stakes deployments (clinical, legal, financial), hallucination mitigation is not optional and must be validated with domain-specific evaluation benchmarks before production.