Feedback Loops Amplifying Bias
回饋迴圈放大偏誤
Continuous Validation
Risk Description
When based on existing data, the system concentrates resources in a particular area, which then generates more records because of the increased activity; due to unmitigated control gaps, these records feed back into the system and further reinforce its original judgment, with the bias visibly widening over several cycles, triggering compliance exposure and operational reputational costs.
Framework Mappings
EU AI ActArt.10、Art.15
NIST AI 600-1Harmful Bias and Homogenization
ISO/IEC TR 24027
ISO/IEC 5338持續確效
MIT AI Risk RepositoryDomain 1
Risk Treatment & Implementation Guidance
Monitor bias-drift indicators tracking whether decision distributions skew across feedback cycles; Isolate or reweight feedback data so system outputs do not directly become the next training basis and self-reinforce; Periodically recalibrate with benchmark data independent of the system's own influence