Stanford Study of 4 Million Applications Shows Hiring Algorithm Audits Can Disguise Bias
The News
A Stanford-led study of 4 million job applications screened by a single vendor's hiring algorithm has found that how an audit is conducted can determine whether bias is detected — or hidden. The research, led by Rishi Bommasani, Sarah Bana, Kathleen Creel, Dan Jurafsky, and Percy Liang, examined outcomes across 156 employers and 11 sectors using Pymetrics' gamified assessment platform.
The study found clear racial disparities in outcomes: 26% of applications from Black applicants and 15% from Asian applicants were directed to positions that adversely impacted them. Critically, Pymetrics' own published audits — which aggregate data across all positions and employers — showed no measurable bias. But when the researchers disaggregated the data position by position, 30% of Black applicants had applied to at least one role where the screening algorithm violated the U.S. four-fifths rule for adverse impact.
The researchers also documented an 'algorithmic monoculture' effect: 10% of candidates who submitted at least four applications were systematically rejected from multiple positions — a rate that significantly exceeds what would be expected under independent hiring decisions. As co-author Kathleen Creel put it: 'We've speculated in past work that if many firms relied on the same AI vendor to screen job applicants, that could prevent some applicants from getting any interviews. But this study was the first time we were able to show this effect in real hiring data.'
Why It Matters
For TA leaders, this study lands like a bucket of cold water. The headline figure — more than 90% of employers use some form of automated screening (World Economic Forum) — is well known. What's new is the demonstration that the standard vendor audit, performed the way most vendors perform it, can miss systemic bias entirely.
If your organization uses an AI screening tool, and you rely on the vendor's published audit, you could be signing off on a process that disproportionately filters out Black and Asian candidates — and you wouldn't know because the audit wasn't designed to catch it. The study also reveals a practical problem for candidates: applying more doesn't help if every employer uses the same vendor's algorithm, because the rejection pattern repeats across companies.
This is not an abstract academic exercise. The Workday litigation — where job applicants allege discrimination by automated hiring tools — is ongoing. Regulators are watching. And as the study's authors note, the fix (per-position auditing with random holdout samples) is straightforward but rarely practiced.