Why You Get Rejected Everywhere: What 4 Million Job Applications Reveal About AI Hiring

You send out application after application and hear nothing back. It is easy to read that silence as a verdict on you. A new study from Stanford suggests something more structural is going on, and it has a name: algorithmic monoculture.

The researchers got access to something almost no one outside the industry ever sees: 4,197,168 real job applications from 3.4 million people, all screened by the algorithms of a single hiring vendor, across 156 employers and 11 industries. Published at the 2026 ACM Conference on Fairness, Accountability, and Transparency (FAccT), it is the first large-scale look at what happens when the same AI stands between millions of applicants and hundreds of employers.

What “algorithmic monoculture” means

Over 90% of US employers now use software to screen or rank applicants, and they buy that software from a handful of vendors. As of 2023, more than 60% of the Fortune 100, and eight of the ten largest US federal agencies, used a single vendor. When one algorithm sits behind many employers, researchers call it a monoculture: lots of separate “decisions” that are really the same decision, made once.

Here is why that matters. When you apply to two companies, you assume you are getting two independent shots. But if the same model scores both applications, you are not. You get one verdict, copied.

Rejections are not independent anymore

This is the study's central finding. Among applicants who applied to four positions, 10% were recommended for rejection at every single one. Among those who applied to ten, 4% were rejected everywhere. That is far more concentrated than chance would predict: a statistical test against a “what if every decision were independent” baseline rejected that idea overwhelmingly.

The most convincing part is the comparison. The authors ran the same analysis on the largest earlier hiring study, 83,000 applications sent to 108 Fortune 500 companies, screened the traditional way by humans and a mix of tools. There, rejections behaved exactly as if each employer decided independently. The correlated, reject-everywhere pattern only appeared under centralized AI screening. In other words, it is not a general feature of job hunting. It is specific to everyone using the same machine.

It does not mean you are unhireable

Now the hopeful part, because the study is careful here. The researchers simulated what would happen if a sample of applicants were scored by every one of the vendor's 495 models. Not a single person was rejected by all of them. The worst-off applicant was still recommended by 52 models, about 11%. Nobody in the data was truly unhireable by the machine.

So systemic rejection is mostly about where you apply, not whether you deserve a job. The people who got rejected everywhere applied to a narrow band of roles scored by correlated models. Apply more widely and your odds change fast, but you have to apply wider than you would think. To push your systemic-rejection risk below 0.1%, the study estimates you would need around 25 applications, versus 10 if every decision were truly independent. The monoculture roughly doubles the volume you need.

The hidden bias in the aggregate

There is a second finding worth knowing. Viewed in aggregate, the vendor's model looked fair: selection rates across racial groups were close enough to pass the US “four-fifths” rule that regulators use. But hiring law is meant to be applied per job, not blended across a whole company. When the researchers broke it down position by position, 10.6% of positions showed adverse impact against Black applicants, and 25.9% of all applications from Black applicants (and 14.7% from Asian applicants) went to positions that disadvantaged them. Aggregation had hidden it. Their recommendation to regulators: measure adverse impact per position, the way the law intends.

What this means for your job search

You cannot see which vendor sits behind a given job, and you cannot change a company's screening. But the study's own findings point to a few things that are in your control.

• Apply wider than feels necessary. The math is unfriendly: plan for volume, because correlated screening quietly raises the number of shots you need.
• Do not be the same input everywhere. The more of your applications a machine sees as one generic profile, the more correlated your outcomes. Tailoring each application to the actual role is how you break that sameness, especially for the resume and text screening most of us face.
• Diversify how you get seen. Referrals, direct outreach, and smaller employers who do not use the big vendors route around the monoculture entirely. One human who reads your application is worth a lot when the alternative is one algorithm reading all of them.
• Stay organized. If you are applying to 25+ roles and tailoring each one, you need a system, or good applications fall through the cracks.

The takeaway

The lesson is not “give up on online applications.” It is that the modern job market has a hidden structure: behind hundreds of employers can sit one algorithm, and that turns your separate applications into correlated bets. You beat correlated bets the same way you beat any of them: place more of them, make each one distinct, and find channels the machine does not control.

Not every rejection is a separate verdict. Sometimes it is the same verdict, echoed. Apply like you know that.

The full study, *Algorithmic Monocultures in Hiring* by Bommasani, Bana, Creel, Jurafsky, and Liang (FAccT 2026), is available on arXiv.