Legal Frameworks and Production Compliance

Legal and regulatory constraints shape fairness implementation across jurisdictions and domains. United States employment and lending laws prohibit disparate impact, creating a legal floor for fairness metrics. European Union regulations add transparency and explainability requirements. Production systems must navigate these frameworks while maintaining business metrics, requiring continuous compliance monitoring integrated into release processes. Disparate impact doctrine, established by Griggs v. Duke Power Co., prohibits practices that disproportionately harm protected classes even without discriminatory intent. The four fifths rule provides a practical test: if the selection rate for a protected group is less than 80% of the rate for the highest group, the practice faces scrutiny. For example, if 60% of Group A applicants are approved, Group B must see at least 48% approval to satisfy the four fifths threshold. Courts consider business necessity as a defense, but the practice must be job related and no less discriminatory alternative can exist. This pushes companies toward 0.8 to 1.25 disparate impact ratio targets in production. Equal Credit Opportunity Act and Fair Housing Act extend these principles to lending, prohibiting discrimination based on race, color, religion, national origin, sex, marital status, age, or receipt of public assistance. These laws create specific compliance obligations: lenders must collect and report protected attributes for monitoring, but cannot use them in credit decisions. This mandates the architectural pattern of sensitive attribute separation, where attributes are collected for auditing but excluded from scoring. Fair lending audits by regulators examine approval rates, interest rates, and terms across demographics, with penalties reaching tens of millions of dollars for violations. General Data Protection Regulation in Europe adds transparency requirements through Article 22, restricting automated decisions that significantly affect individuals without meaningful human involvement. This pushes systems toward explainability and appeal mechanisms. Companies implement model cards documenting training data, fairness metrics, and known limitations, plus human in the loop review for high stakes decisions like loan denials above certain amounts. Production systems combine automated fairness gates with manual audits, processing 95% of decisions automatically while routing 5% to human review based on risk scores or random sampling. Compliance infrastructure integrates into continuous integration and continuous deployment pipelines. Model promotion requires passing fairness unit tests that replay historical data and verify metrics stay within thresholds. Weekly compliance reports go to legal and risk teams, flagging any cohort where confidence intervals approach limits. Incident response playbooks define actions when violations are detected: immediate traffic shift to previous model, root cause analysis within 24 hours, remediation plan within one week. Major platforms treat fairness compliance with the same rigor as security and privacy, with dedicated teams and executive accountability.

Wells Fargo paid $175M in 2012 for fair lending violations after analysis showed Black and Hispanic borrowers paid higher fees than White borrowers with similar credit profiles

Meta implements model cards for all high stakes ML systems, documenting training data demographics, fairness metrics across intersectional groups, and known limitations per GDPR

Google credit model runs daily fairness unit tests on 30 days of historical data, blocks weekly promotion if disparate impact drops below 0.78 or TPR gap exceeds 5.5 percentage points