AI Hallucination
AI hallucination is a common name for an answer with unsupported or false claims. NIST uses the more exact risk term confabulation.
Another example
A legal question asks for a court citation that is not in the approved source set. The system should say the evidence is not enough and ask for or find a source. Clearly labeled fiction or idea-making is not a factual hallucination.
Main parts
- 01generated claim
- 02reference evidence
- 03claim-level mismatch
Use it when
Use this term only after finding a claim that conflicts with trusted facts, source evidence, or stated system limits.
Do not use it when
Find the real cause before using this term. A refusal, style change, old source, search failure, or human error may have another cause.
Name used in code
generated claim ⟂ authoritative evidence Before you ship
On a versioned evaluation set, check the measured baseline, target, and unsupported-claim rate. Verify source traceability, common failures, uncertainty and decline-to-answer behavior, user-visible fallback, human review, monitoring, and rollback.
Outcome: Measure and reduce AI Hallucination in this use case. Do not say it is gone without evaluation evidence. User context: A legal question asks for a court citation that is not in the approved source set. The system should say the evidence is not enough and ask for or find a source. Clearly labeled fiction or idea-making is not a factual hallucination. Failure mode to evaluate: AI Hallucination. Why it applies: Use this term only after finding a claim that conflicts with trusted facts, source evidence, or stated system limits. Evaluation boundary: Find the real cause before using this term. A refusal, style change, old source, search failure, or human error may have another cause. Evaluation requirements: Define the task and source evidence. Use a versioned test set and a clear scoring guide. Track risk groups, unsupported-claim rate, links to source evidence, and when the model must decline to answer. User safeguards: Keep source traceability and state uncertainty clearly. Provide a safe fallback. Require human review for high-impact decisions. Acceptance criteria: Record a measured baseline and target on a versioned evaluation set. Study common failures. Verify fallback, monitoring, and rollback. Evidence and limits (evidence boundary): No single official standard controls the whole term. Its meaning may change across teams, platforms, or frameworks. Unknowns to confirm: Target task, model and version, source data, risk limit, evaluator, and production fallback.
Check this request
Review the current use of AI Hallucination. Definition: AI hallucination is a common name for an answer with unsupported or false claims. NIST uses the more exact risk term confabulation. Release checks: On a versioned evaluation set, check the measured baseline, target, and unsupported-claim rate. Verify source traceability, common failures, uncertainty and decline-to-answer behavior, user-visible fallback, human review, monitoring, and rollback. Before changing code, report the evidence you found, gaps, severity, and the smallest safe fix.