Proxy Discrimination

Proxy Discrimination occurs when an AI system uses variables that appear neutral but correlate with protected characteristics (such as race, gender, or religion), resulting in discriminatory outcomes even without direct use of protected attributes.

Common Proxy Variables:

• ZIP Code/Address: Can correlate with race and socioeconomic status due to historical segregation
• Names: May correlate with ethnicity or gender
• Language Patterns: Can correlate with national origin
• Education Institution: Can correlate with socioeconomic background
• Browser/Device Type: Can correlate with income
• Social Media Connections: Can reflect demographic patterns

Regulatory Context:

• EU AI Act Article 10: Requires examining training data for possible biases that might lead to discrimination, including through proxy variables
• GDPR: Special categories of data include not just explicit sensitive data but data that could be used to infer such characteristics
• US Guidance: EEOC guidance notes that discrimination can occur through "facially neutral" factors

Why Proxy Discrimination is Challenging:

• Proxies can emerge from complex feature interactions
• May not be intentionally designed
• Difficult to detect without careful analysis
• Removing one proxy may cause the model to rely on others
• Some correlated features may be legitimately predictive

Detection Methods:

• Feature importance analysis
• Counterfactual fairness testing
• Subgroup analysis of outcomes
• Correlation analysis with protected attributes

Mitigation Approaches:

• Pre-processing to remove proxy signals
• Fairness constraints during training
• Post-processing adjustments
• Regular auditing and monitoring