hai.stanford.edu web signal

Stanford HAI Study Finds Racial Bias in AI Hiring Screeners

TL;DR

  • A Stanford HAI study of 4 million applications found 26% of Black applicants faced AI-screened positions that discriminated against their racial group.
  • Roughly 40,000 more applications would have advanced if the AI tool had recommended candidates at equal rates across racial groups.
  • Ten percent of applicants who submitted four applications screened by the same vendor were rejected from all of them, exceeding independent-probability expectations.

The concern about AI hiring tools and bias has circulated long enough to start feeling like a theoretical placeholder. Stanford HAI just made it empirical.

A research team including Rishi Bommasani, Sarah H. Bana, Kathleen A. Creel, Dan Jurafsky, and Percy Liang examined 4 million job applications across 1,700 positions from 150 employers in 11 industry sectors, all screened by a single third-party AI vendor. Applying the EEOC's four-fifths rule from Title VII of the Civil Rights Act, they found that 26% of Black applicants and 15% of Asian applicants encountered positions where the AI system screened their group at a discriminatory rate. If candidates had been recommended at equal rates across groups, roughly 40,000 additional applications would have advanced.

The second finding is arguably the more structural one: what the researchers call "algorithmic monocultures." When an applicant submitted multiple applications handled by the same vendor's tool, rejection rates exceeded what you would expect if each decision were statistically independent. Ten percent of applicants who submitted four applications were rejected from every single one. That pattern did not appear in historical hiring data from Fortune 500 companies that used traditional, non-AI screening methods, which points specifically at market concentration of AI tools rather than applicant quality as the driver.

What the study does not give you is the name of the vendor, so it is impossible to say whether these are failures specific to one system's design or a generalizable property of how resume-screening AI is trained and deployed at scale. The mechanistic source of the bias, whether training data, feature selection, or model architecture, is also not addressed.

For employers, the practical implication is liability: the four-fifths rule is enforceable, and these findings suggest many companies may be running tools that would fail a regulatory audit without knowing it. For vendors competing in hiring AI, rigorous independent auditing on equity metrics is now a differentiator the market has empirical reason to demand.

Shared on Bluesky by 3 AI experts