Hallucination

What is Hallucination?

Hallucinations are reduced — not eliminated — by good retrieval, citations, and evaluation. In an enterprise setting, the more dangerous failure mode is a hallucinated action, not a hallucinated answer: an agent that fabricates a justification for the wrong tool call. See Private RAG vs Enterprise Search and Agent Evaluation.

What is an example of Hallucination?

An assistant cites a policy section that does not exist and invents an eligibility rule. A grounded system should retrieve the authoritative policy, quote the relevant passage, provide a resolvable citation, or abstain when evidence is insufficient.

How is Hallucination different from related concepts?

A hallucination is an ungrounded model output. Stale information may be faithfully generated from outdated data, and a retrieval error may supply the wrong evidence; both can cause incorrect answers through different mechanisms.

What should enterprises evaluate for Hallucination?

• Build evaluation sets with answerable, ambiguous, adversarial, and intentionally unanswerable questions.
• Measure claim-level support and citation correctness rather than scoring only whether an answer sounds relevant.
• Require abstention, verification, or human review when evidence is weak or consequences are material.

Related terms

• RAG (Retrieval-Augmented Generation)
• AI Guardrails