Harm Difficult to Perceive or Quantify
傷害難以感知與量化
Continuous Validation
Risk Description
When the system's determinations produce long-term adverse effects on a particular group, yet each individual decision appears reasonable; due to unmitigated control gaps, affected individuals have no way to connect their situation to the system, and the organization, lacking any discrete incident, never initiates a review, triggering compliance exposure and operational reputational costs.
Framework Mappings
EU AI ActArt.26、Art.72
ISO/IEC 42001Annex A.6.2.6、A.5.4–A.5.5
NIST AI RMFMEASURE 2.11、MANAGE 4.1
ISO/IEC 5338運作與監控
MIT AI Risk RepositoryDomain 1、Domain 6
Risk Treatment & Implementation Guidance
Establish longitudinal statistical monitoring of system decisions by affected group and periodically review for systematic adverse trends; Define quantifiable harm indicators with thresholds that trigger review procedures even in the absence of a discrete incident