Societal Bias Embedded in Historical Data
歷史資料中的社會偏誤
Design & Development
Risk Description
When an organization uses automated screening models for recruitment or resource allocation; because the training dataset inherits historical societal bias without de-biasing preprocessing, the model converts neutral features into proxy discriminatory variables, systematically denying equal opportunity to minority groups and exposing the firm to employment discrimination lawsuits and ESG rating downgrades.
Framework Mappings
EU AI ActArt.10
NIST AI 600-1Harmful Bias and Homogenization
NIST AI RMFMEASURE 2.11
ISO/IEC TR 24027
ISO/IEC 5338設計與開發
MIT AI Risk RepositoryDomain 1
Risk Treatment & Implementation Guidance
Audit datasets for bias before training, examining group distributions and label tendencies in historical records; Monitor cross-group fairness metrics—pass rates and score distributions—before and after launch; Retain scoring-basis records for personnel systems so appeals can be explained and re-reviewed