A new study by researchers at the Stanford Digital Economy Lab and Stanford HAI analyzed 3.4 million job seekers and 4 million applications screened by a single third-party vendor's AI hiring tool, revealing racial disparities in candidate selection. The tool, which uses game-based assessments to generate "recommend" or "do not recommend" labels, produced adverse impact-defined by the EEOC's four-fifths rule-for 26% of Black applicants and 15% of Asian applicants. The authors estimate that roughly 40,000 more applications from those groups would have advanced if recommendation rates matched those of the most-favored demographic.
How the vendor's algorithms were tested
The study tracked individuals across 1,700 job postings at roughly 150 employers, all using the same vendor's screening technology. Reporting identifies the vendor as pymetrics and its parent company Harver. The platform administers game-based assessments and outputs binary recommendation labels that employers incorporate into hiring workflows. The research team applied position-level adverse-impact calculations using the EEOC's four-fifths rule, which flags potential discrimination when the selection rate for a protected group falls below 80 percent of the highest group's rate.
In contrast, the vendor's previous analyses pooled recommendations differently and concluded no widespread adverse impact. That methodological gap is the core of the paper's argument: aggregate checks can mask legally relevant disparities that surface only when data is broken down by position. The findings were covered by Fortune, which reported that Harver did not respond to a request for comment. The paper will be presented at an academic venue.
Algorithmic monoculture and systemic risk
The researchers also examined what happens when similar screening algorithms are deployed broadly across industries. They call the resulting pattern algorithmic monoculture, where a few vendors or design approaches dominate hiring pipelines. Because the algorithms are trained or validated on comparable data, they can produce correlated errors. This creates a systemic risk: a candidate rejected by one employer's tool may face the same outcome from another company using a similar model.
The study documents that employers in finance, manufacturing, and technology sectors rely heavily on such third-party screening. The concentration amplifies bias beyond individual firms and makes it harder for affected individuals to find alternative pathways.
Why this matters for scientists and researchers
Data scientists and ML engineers auditing or building hiring tools need position-level fairness metrics that match legal standards like the four-fifths rule. Aggregate tests, which look at overall recommendation rates without slicing by job opening, can obscure adverse impact that emerges at the position level. The study also shows that audit signals can propagate: models that appear similar in structure or training data can introduce correlated failures when deployed across many employers.
Understanding these legal metrics requires not only technical skill but also domain knowledge; resources like AI for Human Resources can help data scientists align fairness testing with employment law. For researchers, the paper offers a blueprint for replicable audits that use legally grounded metrics to measure discrimination in AI screening tools.
Your membership also unlocks:
