Stanford Study Finds Racial Bias in AI Hiring Algorithm Affecting Black and Asian Applicants
A Stanford study found that a hiring algorithm used by pymetrics (now owned by Harver) disproportionately screened out Asian and Black job applicants across multiple positions. Earlier audits failed to detect the bias because they analyzed hiring outcomes in aggregate rather than by individual job, and applicants rejected by one employer were more likely to be rejected by others using the same system.
What happened
A Stanford-led study found that a hiring algorithm administered through pymetrics, now owned by Harver, disproportionately screened out Asian and Black job applicants. The study examined assessments across roughly 29,000 applications and found that candidates of Asian descent faced the largest cumulative shortfall in recommendations among racial groups analyzed, with approximately 29,000 more Asian applications that would have advanced if they had been selected at the same rate as the most-selected racial group for each position.
A position-by-position review found adverse impact affecting Asian and Black applicants across multiple jobs using the four-fifths rule, a common employment discrimination benchmark. Earlier audits had not detected these racial disparities because they evaluated hiring outcomes in aggregate rather than by individual job, allowing harms at specific positions to disappear when results were combined into a single analysis.
The study identified a pattern called 'systemic rejection' in which applicants screened out by one employer were more likely to be rejected by others using the same technology. Among applicants who submitted four applications through employers using the platform, 10% were rejected from every position, a rate that exceeded what would be expected if each employer had made independent decisions.
Reported impact
Asian and Black job applicants were disproportionately filtered out of the hiring process across multiple positions. The findings challenged assumptions that removing resumes and demographic information automatically reduces bias, as the behavioral games assessing traits such as attention, risk tolerance and decision-making could still correlate with race and produce unequal outcomes.
Index analysis
The study argues that employers and regulators risk overlooking discriminatory outcomes when they rely on broad audits rather than examining how algorithms perform in individual jobs. The characteristics measured by the assessments could correlate with protected characteristics, producing unequal outcomes even when demographic information was not explicitly used in the system.
Governance controls
Earlier bias audits of the system failed to detect significant racial disparities by evaluating hiring outcomes in aggregate rather than by individual job, demonstrating gaps in existing audit methodologies.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.