Glossary
hallucination
A model output that is fluent and plausible-sounding but factually wrong, ranging from invented citations and APIs to fabricated names, dates, and quotes.
A model generates a confident assertion that sounds right but is false. The pattern is not “the model does not know,” it is “the model generates text that pattern-matches knowledge without being grounded.” Classic examples: invented academic citations, fabricated API method names, hallucinated court cases that lawyers have cited and been sanctioned for, and convincing-sounding biographical detail about real people.
Mitigations stack at different layers. RAGretrieval-memoryA pattern where a model retrieves relevant documents from an external store at query time and conditions its answer on them, instead of relying only on parametric knowledge. Open full entry grounds answers in fetched documents. Tool calls verify facts against authoritative sources. Self-consistency sampling reduces some classes of error. Refusal training teaches the model to decline rather than fabricate when uncertain. Reasoning-mode models verify intermediate steps. None of these eliminates hallucination; together they reduce the rate.
For verification systems (citation linters, fact-check pipelines, the audit work in this site), hallucination is the failure mode the infrastructure exists to catch. The cost of one hallucinated number in authoritative-looking prose is much higher than the cost of saying “I do not know.”