Algorithmic Monocultures in Hiring Amplify Position-Specific Racial Disparities

Over 90% of U.S. employers rely on hiring algorithms from a handful of vendors, creating monocultures that produce homogeneous rejections exceeding independent baselines in analysis of 3.4 million applicants. The github.io study of 4 million applications across 156 employers found 25.87% of Black-submitted applications and 14.74% of Asian-submitted applications directed to positions showing adverse impact under Title VII four-fifths rule, with Black applicants facing impact in 30% of cases when disaggregated by position. Aggregate vendor-wide metrics previously masked these effects, while shortfall calculations indicate 29,000 additional Asian applications would qualify if selected at top-group rates. Systemic rejection rates surpass statistical independence predictions, confirming theoretical monoculture risks in deployed systems. Prior aggregate studies missed per-position signals required by U.S. employment law, and data-access barriers have blocked independent verification despite Fortune 100 adoption rates above 60% for single vendors like HireVue. Related empirical work on fairness constraints shows similar convergence effects in other high-stakes domains when models share training pipelines.