Hallucination
FoundationsGovernance, Safety and Ethics
A hallucination is an AI output that states false information with complete confidence: invented citations, fabricated statistics, fictional case law. The term is contested. Some researchers argue it flatters the model by implying something dreamlike, when the accurate description is simply a factual error.
In practice
AI output needs human verification wherever a confident mistake carries regulatory or financial consequences, which covers most of the uses a mid-market business would actually want. Design the verification step into the pilot, because a system that is right most of the time is the hardest kind to police.