When one AI vendor screens hiring for a whole industry, the same people get rejected everywhere

One algorithm, 150 employers, 4 million applications

Stanford HAI's writeup of the Algorithmic Monocultures in Hiring paper follows 3.4 million applicants submitting 4 million applications to 1,700 postings across 150 employers and 11 industry sectors, every one screened by the same third-party vendor, pymetrics, whose tool has applicants play assessment games rather than read resumes off the page.

Two findings sit on top of each other. The first is disparate impact. Applying the EEOC's four-fifths rule (which flags a position when one group is recommended at under 80% of the most-recommended group's rate), the researchers report 26% of Black applicants and 15% of Asian applicants hit positions showing adverse impact against them. Their headline counterfactual: if the system had recommended Black and Asian candidates at the rate of the most-favored group, 40,000 more applications would have advanced.

The second finding is the one the article says is easy to miss: systemic rejection. Because a single vendor mediates screening for many employers, an applicant rejected once tends to be rejected everywhere. Ten percent of people who submitted four applications were rejected from all of them — far above what you'd expect if each employer decided independently. The researchers checked this against the largest prior study of first-round screening (83,000 synthetic resumes sent to 108 Fortune 500 firms, no single AI gatekeeper) and found that there, systemic rejection sat right at the independent-decision baseline. The correlation is specific to the monoculture.

The thread fights over what "adverse impact" even proves

The comments split, predictably, on whether a disparate outcome is evidence of bias at all. That argument is old; the more useful contributions were procedural and statistical.

Several readers misread the four-fifths rule as a claim of proven discrimination. One commenter who'd clearly worked with it corrected the record repeatedly:

"The four fifths rule is used to define if there is a 'substantially different selection rate'. It does not measure racial discrimination… It's purely a signal for 'we should consider asking more questions, because this appears unusual'." — a commenter who walked through the EEOC standard

Another located the framework in employment law, noting this is the disparate-impact doctrine that pushed employers off IQ-style screening tests decades ago, and that the common objection is precisely that it can produce unfalsifiable claims. — a commenter on the disparate-impact doctrine

The sharpest reframing argued the bias question is almost a distraction from the structural one:

"even a small bias, if used by virtually everyone, can be damning… If we move to using just a small number of AI models to help do things like hiring, we will amplify biases and possibly completely lock out portions of the population." — a commenter who reframed it as a concentration risk

One reply pushed that further: you don't even need bias for the monoculture to hurt. If every employer shares one fuzzy ranking instead of many independent ones, the line between hireable and unhireable gets much sharper — a point another commenter connected to Simpson's paradox, since the adverse impact vanishes when you pool all of pymetrics' recommendations and only appears position by position.

The skeptics weren't quiet either. Multiple commenters flagged that the writeup's title says AI when the underlying tool is a psychometric game, that the study leans on the four-fifths flag rather than controlled causal analysis, and that "a rejected resume is more likely to be rejected by every other employer" is just what you'd expect from real signal, not bias. The paper's own answer is the comparison dataset: real signal alone produced no excess systemic rejection in the non-algorithmic Fortune 500 study, and that's the gap the monoculture opens.