When the Same AI Rejects You Everywhere
Over 90% of U.S. employers use algorithms to screen job applicants. Over 60% of the Fortune 100 use HireVue. These aren’t competing systems — they’re the same algorithm, from the same vendor, deployed across hundreds of companies.
That means the same AI that decided you weren’t right for Company A just made the same call about Company B. And Company C. And Company D.
A new study published at FAccT 2026 by Bommasani, Bana, Creel, Jurafsky, and Liang from Stanford, Chapman, and Northeastern gave us the first large-scale look at what this actually does to people. The dataset: 3.4 million real job applicants submitting 4 million applications to 156 employers across 11 market sectors. Every single application was assessed by algorithms from a single vendor.
The findings are the kind that should make C-suites nervous.
Twenty-five-point-eight-seven percent of applications submitted by Black applicants went to positions where the algorithm showed adverse impact under Title VII of the Civil Rights Act. For Asian applicants: 14.74%. Previous studies missed this because they averaged across all positions instead of looking at each one separately, which is how Title VII requires it be measured. The aggregate smooths out the harm. Disaggregate, and the harm is right there.
Then there’s the systemic rejection problem. Among applicants who submitted 4 applications, 10% were rejected from every single one. The researchers compared this to a baseline of independent decisions (where each employer evaluates you on their own, without a shared algorithm) and found the observed rate of total rejection was massively higher — with a chi-squared of 18,481 and p < 0.001. In plain English: this isn’t bad luck. It’s the algorithm.
The researchers contextualized this using data from Kline et al. (2022), which sent 83,000 fake resumes to 108 Fortune 500 firms before AI screening was this pervasive. In that data, systemic rejection rates matched the baseline of independent decisions almost perfectly (p = 0.69). The excess homogeneity — the structural “no” — only appears when a central algorithm is making the call.
This is what an algorithmic monoculture looks like. It’s not just that one AI has bias. It’s that the same biases replicate across the entire job market simultaneously. You don’t get rejected by one company and move on. You get rejected by the system.
And here’s the part that really gets me: this is the only independent research group that managed to get access to this data. Paragraph four of their findings is blunt about it: “Data access inhibits independent research into hiring algorithms.” Policy intervention may be necessary, they say. That’s academic code for “we couldn’t get the data without a lawsuit, and neither can anyone else.”
Think about that. The system that decides who gets a shot at a job — arguably the most consequential gate in modern life — is a black box that only one academic team has been allowed to inspect. And what they found is that it systematically funnels disadvantage along racial lines and locks people out of the market wholesale.
Monocultures are fragile. Ask a farmer. Ask a sysadmin who watched one zero-day take down every server running the same software. Every field learns this lesson eventually: diversity isn’t a luxury, it’s resilience. Hiring should learn it too, before the algorithm decides that lesson doesn’t apply.
Sources: Algorithmic Monocultures in Hiring (arXiv:2605.27371) by Bommasani et al., published at FAccT 2026; project site