Hallucination

Hallucinations happen when an LLM generates information that wasn't in its training data, isn't supported by retrieved context, and isn't true — but is delivered with the same confidence as accurate output. The model doesn't know what it doesn't know.

Common hallucination patterns: invented citations, invented function names or APIs, invented quotes, plausible-sounding but wrong factual claims (especially about niche topics, recent events, or specific numbers).

Mitigations operators should know: ground responses in retrieved sources (RAG), require citations, use evals to catch hallucination patterns specific to your domain, and design UIs that don't pretend the AI is always right. The most-mature AI products show their work — "here's the source I'm pulling from" — rather than just answering.