AI hiring tool shows bias in 11% of individual jobs

What happened

Researchers analyzed 4.2 million applications to pymetrics, a game-based hiring tool used by 156 employers. When they split results by individual job position (1,746 total), roughly 11% of jobs showed adverse impact against Black applicants—a bias that was hidden when results were pooled company-wide. The tool trains on current employees as 'good' examples and random people as 'bad,' so it learns who resembles existing staff rather than who can perform the job.

Why it matters

The study confirms a real compliance problem: company-level fairness audits can mask discrimination in individual roles. However, the viral panic about 'algorithmic monoculture' overstates the actual risk. The researchers' own data shows 84% of applicants applied to exactly one position, and only 0.02% applied to ten positions—the scenario required for the 'rejected everywhere' nightmare hardly occurs in practice.

What to watch

The headline finding worth keeping is that if you audit hiring only at the company level, you should change it to look at individual jobs. The broader 'monoculture' concept the paper leads with (citing HireVue's use by over 60% of Fortune 100 companies) describes a different product category; the study itself examined only pymetrics, a smaller competitor using games rather than the structured-interview methods HireVue deploys.