AI Glossary / Hallucination

🛡️ Safety and quality

Hallucination is when a model confidently states falsehoods: an invented date, a nonexistent book, a fabricated quote or legal provision. It's not "lying" - the model has no concept of truth, it generates statistically probable text, and probable can be false. Hallucinations are the main weakness of language models in factual applications. Grounding and RAG, search modes, and asking for sources reduce them - but no technique eliminates them entirely. Practical rule: the more important the decision, the more thoroughly you verify the answer.