Sendhil Mullainathan, "Discrimination by Algorithm and People" | University of Chicago Law School

algorithmic-biasfairnessmlcriminal-justicehealthcarediscrimination

Abstraction: Algorithms can reduce discrimination but require explicit equity objectives and correct labels

Key points:

Resume audit study: white names received callbacks 50% more than identical resumes with Black names (8.5% vs 6.19%); gap widens for higher-skill resumes
NYC pretrial release algorithm simulation: same crime rate achievable with 40% fewer people jailed, or same jail population with 25% less crime — equivalent to closing Rikers Island
Healthcare care coordination algorithm (applied to 120M+ people) enrolled Blacks at 17% vs. 36% appropriate rate due to optimizing cost as proxy for health; Blacks are sicker at every cost risk-score level
Root cause: label misspecification — cost was used as a proxy for health need; Blacks with same sickness cost less due to lower healthcare access
Algorithm vendor achieved 84% reduction in racial disparity gap after being contacted, supporting "misunderstanding" over malice as explanation
Three lessons: (1) equity must be built into the objective, not a collateral concern; (2) label misspecification is a critical source of bias; (3) algorithms are far easier to audit, diagnose, and fix than human decision-making

Connections: Sendhil Mullainathan · Algorithmic Bias · Fairness In ML · Predictive Policing · Label Bias