A Stanford study finds one AI hiring tool quietly screened out Black and Asian applicants

A team at Stanford HAI looked at 4 million job applications from 3.4 million candidates, across 1,700 postings at 150 employers in 11 industries, all screened by one dominant third-party vendor's machine learning model. Using the EEOC's four-fifths rule as the test, they found that 26 percent of Black applicants and 15 percent of Asian applicants applied to positions where the system discriminated against their racial group. Had the tool recommended candidates evenly, roughly 40,000 more applications would have moved forward.

The methodological point is as important as the headline. When the researchers pooled all of the vendor's recommendations together, the bias disappeared from view. It only became visible when they examined each job posting on its own. That is a warning for anyone who audits these systems by looking at aggregate pass rates, because the aggregate can look clean while individual roles are skewed. The study also surfaced a second problem the authors call systemic rejection: because so many employers use the same vendor, an applicant rejected once tends to be rejected everywhere. About 10 percent of people who submit four applications are turned down by all of them.

That concentration is the structural issue. When one model sits behind a large share of the hiring market, its blind spots are not isolated mistakes, they are gates that close across many employers at once.

Why it matters

If you run hiring and rely on a screening vendor, this says your fairness checks need to be per-role, not pooled, or you will miss exactly the bias regulators care about. For job seekers, it explains a real pattern: repeated automated rejections may reflect one shared model's blind spot rather than your qualifications.