SystemOverflow
Learn→Privacy & Fairness in ML→Bias Detection & Mitigation→6 of 6
Privacy & Fairness in ML • Bias Detection & MitigationHard⏱️ ~3 min

Legal Frameworks and Production Compliance

Key Regulations

Equal Credit Opportunity Act (ECOA): Prohibits discrimination in credit decisions based on race, color, religion, national origin, sex, marital status, age. Applies to any model used in lending. Fair Housing Act: Prohibits discrimination in housing-related decisions. Advertising algorithms showing housing ads to specific demographics can violate this. Title VII: Prohibits employment discrimination. Hiring algorithms must not have disparate impact on protected groups. GDPR Article 22: Gives EU citizens right to explanation for automated decisions. ML models must provide meaningful information about decision logic.

The 80% Rule (Disparate Impact)

The EEOC 80% rule: selection rate for any group should be at least 80% of the rate for the highest group. If 50% of Group A applicants are hired and only 30% of Group B, the ratio is 30/50 = 0.6, violating the threshold. This is not a safe harbor: passing does not guarantee compliance, failing does not guarantee violation. But it is the primary statistical test regulators use. Document your demographic parity ratio for every model touching regulated decisions.

Compliance Architecture

Audit trail: Log every prediction with features and outcome. Store for 5 years minimum for lending, 3 years for employment. Model documentation: Maintain model cards documenting training data demographics, fairness metrics, intended use. Adverse action notices: Credit denials must explain reasons. ML models need interpretability to generate specific reasons ("insufficient income" not "low probability score"). Testing protocol: Regular fairness audits by independent team. Pre-deployment fairness certification for high-risk models.

⚠️ Key Trade-off: Compliance requirements vary by jurisdiction and domain. Employment, lending, and healthcare have different rules. Consult legal counsel for specific obligations.
💡 Key Takeaways
✓Key regulations: ECOA (credit), Fair Housing Act, Title VII (employment), GDPR Article 22 (explanation rights)
✓EEOC 80% rule: group selection rate should be at least 80% of highest group rate
✓Audit trails: 5 years for lending, 3 years for employment decisions
✓Model cards document training demographics, fairness metrics, intended use
✓Adverse action notices require interpretable reasons, not just probability scores
📌 Interview Tips
1Explain 80% rule with concrete calculation: 30%/50% = 0.6 violates threshold
2Mention compliance varies by domain: employment, lending, healthcare have different rules
← Back to Bias Detection & Mitigation Overview