When One Company's Hiring Algorithm Decides for Millions of Job Seekers

A new study from Northeastern and Stanford universities looks at what happens when the same automated hiring screener is used to evaluate millions of job applications across many companies. The findings show that relying on a single vendor's system can create serious, hidden problems — and the researchers found evidence of unfair outcomes that may already be affecting real job seekers today.

The Basic Problem

Think of how a hiring team might work the old way: multiple people screen applications, each one bringing their own judgment. When one person makes a mistake or has a blind spot, another might catch it. The errors are spread around.

Now imagine the opposite. One computer algorithm does all the screening for all the positions at a company — or, at scale, for many companies using the same vendor. If that algorithm has a flaw, that same flaw gets applied the exact same way to every single application it sees. There is no human variation to balance it out. This situation — where one system makes the same decision the same way for millions of people — is what the researchers call an "algorithmic monoculture."

The researchers studied a real dataset of 3 million job applicants and 4 million applications, all screened by the same vendor's algorithm. This is not a theoretical problem. This is how many large hiring systems actually work right now.

What They Found

The study looked at people who applied for 10 or more jobs. Among them, 4% got rejected for every single position they applied for. On its own, that might not sound strange — some people are genuinely not a good fit. But here is what matters: this 4% rate is higher than it should be by random chance. The algorithm was not just making independent mistakes. It was making the same mistakes over and over.

For a job applicant, this means rejection after rejection, but they would never know why. From their perspective, it just looks like bad luck. In reality, a single computer program was rating them low for every job, regardless of whether their skills matched each position or not.

Outcomes Differed by Race

The researchers also looked at whether the algorithm affected different racial groups differently. Using the legal standard that U.S. employers already have to follow — a rule that asks whether a hiring tool produces significantly lower pass rates for one group than another — they found this:

For Asian applicants, 14.74% of the applications went to jobs where the algorithm created this kind of disparity. For Black applicants, the number was 25.87% — roughly one in four applications.

These numbers matter because they reflect real conditions in hiring systems that companies are using today. One-quarter of applications from a group of job seekers being caught in a potentially unfair screening system is not a small edge case.

Why This Matters Beyond Hiring

We have seen something similar happen before, in different parts of technology. In the 2010s, when Google, Facebook, and a few other platforms became the main way content reached people online, the diversity of what people saw got narrower. One company's algorithm set the rules for everyone. Over time, this changed how information spread and who could build an audience.

The hiring problem works the same way structurally: a single decision-making system gets concentrated in the hands of one vendor, and suddenly the outcomes for millions of people become correlated and systematic rather than random and local. But employment decisions are subject to existing anti-discrimination laws in ways that social media platforms are not. This means the legal risk for companies using these systems is clearer and more concrete than some HR departments may realize.

What Comes Next

For companies deciding whether to buy or build hiring algorithms, this research points to some practical problems.

First: relying on a single vendor creates risk. If that vendor's system has bias or flaws, those problems spread across every company using it at the same time. No single employer can contain the problem on their own.

Second: the kinds of algorithms used in hiring are deterministic, meaning they produce the exact same result every time they see the same person. That is what makes them predictable and auditable. But it also means if there is bias baked in, that same bias gets replicated identically, thousands or millions of times, until someone fixes the underlying system.

Third: new laws are coming that will require companies to audit their automated hiring decisions. The European Union's AI Act and similar laws in several U.S. states will mandate this kind of scrutiny. This research shows what that scrutiny should look like and what kinds of problems it can find.

The researchers do not propose specific solutions in this paper — their job was to diagnose the problem clearly. But the diagnosis is detailed and grounded in real data. This is not a hypothetical concern. This is how hiring actually works for millions of people right now.

Solutions exist — using multiple vendors rather than one, having humans review borderline cases, regularly checking whether the system treats different groups fairly. None of these require new technology. The gap has been organizational and business decision-making, not technical capability. Research like this one makes it harder to ignore that gap.