← Glossary/Governance & risk

Defined term

Hallucination

Plausible but factually incorrect output generated by an LLM with no grounding.

Hallucination is when a model produces an answer that sounds confident but is not supported by the provided sources or the real world. Hallucination cannot be eliminated by prompt alone; production defenses combine grounding (RAG), confidence scoring, citation requirements, output validation, and human review for low-confidence cases.

When it matters

Highest-stakes failure mode in any AI workflow. A hallucinated number, citation, or fact can ship a wrong contract, a wrong dose, a wrong claim decision. Defense-in-depth required: grounding + validation + reviewer queues.

Real example

A model citing 'NIST SP 800-53 control AC-22' when the document was about AC-21 — a single-digit hallucination that would have passed a tired reviewer. Caught by a citation-validator that re-fetched the source and flagged the mismatch.

KPIs to watch

Hallucination rate on factual claims (<1% target for production), citation-validator catch rate (>95%), reviewer override rate trending up (signal that hallucinations are slipping through).

Related terms

See it in action

We use this every week

Book a 30-min call and we'll walk you through how Hallucination shows up in a real engagement we're running.

Book a 30-min call