Indirect Discrimination Through Proxy Variables
代理變數造成間接歧視
Verification & Validation
Risk Description
When an organization deliberately excludes protected attributes from model inputs, believing fairness is assured; due to unmitigated control gaps, subsequent analysis finds the model reconstructs the same distinction through variables such as place of residence, producing outcomes no different from using the attribute directly, triggering compliance exposure and operational reputational costs.
Framework Mappings
EU AI ActArt.10
NIST AI 600-1Harmful Bias and Homogenization
NIST AI RMFMEASURE 2.11
ISO/IEC TR 24027
MAESTRO
ISO/IEC 5338驗證與確效
MIT AI Risk RepositoryDomain 1
Risk Treatment & Implementation Guidance
Analyze proxy variables, testing correlation structures between non-sensitive features and protected attributes; Verify outcome-level equivalent discrimination with causal fairness testing rather than input exclusion alone; Neutralize confirmed proxies and monitor outcome fairness continuously