S0064RS2-T05-S0064-Z · Full risk code

Confidence Miscalibration

信心度校準失準

Verification & Validation
Risk Description

When an organization designs an automation threshold based on confidence scores, releasing high-confidence cases automatically; due to unmitigated control gaps, the system consistently assigns high confidence to a particular class of errors, so those cases all bypass human review and are discovered only after losses accumulate, triggering compliance exposure and operational reputational costs.

Framework Mappings

NIST AI 600-1Confabulation
OWASP Top 10 for LLMLLM09
NIST AI RMFMEASURE 2.5
ISO/IEC 5338驗證與確效
MIT AI Risk RepositoryDomain 7

Risk Treatment & Implementation Guidance

Validate confidence-score calibration before launch, confirming high confidence corresponds to high accuracy before using it as an automation threshold; Retain sampled human review of high-confidence auto-approved cases to detect systematic overconfident errors; Apply uncertainty quantification, forcing out-of-distribution inputs to human handling