Overview
- A Stanford-led team analyzed about 4 million applications from roughly 3 million applicants to 1,746 roles at 156 employers and found 10.62% of jobs produced adverse impact against Black applicants under the EEOC four-fifths rule.
- The researchers report that 25.87% of applications by Black job seekers—nearly 40,000 submissions—were for positions where the algorithm’s outcomes meet federal guidelines for discrimination.
- The paper documents an 'algorithmic blackball' where correlated vendor models cause repeat rejections: among people who applied to 10 roles screened by the same vendor, 4% were rejected from every one and Pymetrics can reuse scores for up to 330 days.
- The team showed that pooling results across employers or occupations can hide harms, and that position-by-position auditing using the EEOC four-fifths test uncovers disparities that vendor-level aggregates miss.
- The authors will present the findings at the ACM FAccT conference and urge policy steps including position-level audits, cross-employer market surveillance, limits on vendor concentration, and legal pathways for independent researcher access as the EU AI Act compliance date approaches.