Protected Class Proxy Discrimination

Issue: Resume screener learns that names, dates, or graduation years correlate with performance; uses these as hidden proxy signals for protected characteristics (age, race, national origin)

Frequency: Common

Symptoms

• AI rejects candidates with “dated” graduation years without mentioning age
• Candidates with ethnic-sounding names rejected at higher rate
• Job history gaps (maternity leave, visa sponsorship) correlated with rejection
• EEOC audit flags disparate impact: different acceptance rates by demographic

Root Cause Models optimize for “success prediction” based on training data that reflects historical biases. If hiring data shows older workers were hired less, model learns to downweight resume signals correlated with age (graduation year, long tenure). This is illegal discrimination disguised as predictive hiring.

Example

Training data: 1000 previous hires with performance ratings
Historical pattern: Workers hired >15 years ago (older candidates) avg rated 3/5
Younger hires (grad year 2015+): avg rated 4/5
Model learns: Graduation year predicts performance
Decision rule learned: Year 2000-2005 graduates → lower score
Resume: "Graduated 2003; VP Engineering at 3 companies; proven track record"
Model decision: REJECT (graduation year 2003 = proxy for "too old")
EEOC audit: Disparate impact against >40 age group (95% CI significant)
Violation: Age discrimination under ADEA

Key Statistics

Mitigation Strategies

1. Remove protected proxies: Delete graduation year, address, gaps from input features
2. Fairness testing: Measure accept rate by protected class; ensure statistical parity
3. Bias audit: Test model against synthetically modified resumes (same qualifications, different names/dates)

Production Signals

Alerts

References

• AI Hiring Discrimination - Research on proxy discrimination
• EEOC AI Guidance - Regulatory framework