Hallucination

An AI hallucination is when a language model confidently produces text that sounds plausible but is factually wrong — wrong dates, fake citations, invented quotes, fabricated case law. It happens because LLMs predict what should come next based on patterns, not what is actually true. When the model has never seen specific information, it generates the most likely-looking continuation, which can be invented from whole cloth. In 2026, hallucination rates have dropped significantly (GPT-5 hallucinates roughly half as often as GPT-4 did) but the problem is not solved. The only reliable defences are: verify specific factual claims yourself, use RAG (Retrieval Augmented Generation) so the model is reading real source documents, and stay skeptical of confident-sounding but unverified claims.