The concept
Useful uncertainty is actionable. An answer should distinguish what is supported, what is inferred, and what information is missing. A confident tone is not a measurement of correctness. Design a path for deferral: ask for a missing document, flag a disputed fact, or send the case to a qualified reviewer.
Worked example
A fictional historical society asks an assistant to identify a person in an undated photograph. The archive contains a similar portrait but no matching catalog record. A useful response lists visible details and requests provenance; an unsafe response confidently supplies a name. The team creates an 'insufficient evidence' outcome and measures whether reviewers receive the missing information needed to resolve the case, rather than treating every refusal as a failure.
Limits and judgment
Abstention has costs. Excessive deferral can overload reviewers and make a tool useless, while vague caveats can disguise unreliable answers. Evaluate whether the system declines for the right reasons and whether users understand the next step. Where possible, tie uncertainty to a specific gap in evidence instead of an unsupported numerical confidence score.
FROM UNDERSTANDING TO ACTION
Put it to work.
Rewrite one overconfident claim into an evidence-based answer with a concrete next step.
RETRIEVE. REASON. REMEMBER.
Knowledge check.
Choose one answer for each question, then check your reasoning.
Sources & further reading.
Links support the core concepts or identify the book. Examples, exercises, and critical analysis are original Institute material. Source checks are not an independent expert review of every claim.
- NIST AI Risk Management Framework ↗
Core concept; scenario and analysis are original educational examples. Source reviewed 2026-10-19.