Your Hiring-AI Risk Isn't Independent: Algorithmic Monoculture in Recruiting

Ten percent of applicants who submit four applications get rejected by every single one. The cause runs deeper than being unqualified for all four: the four employers are running variations of the same screening algorithm, and that algorithm reached the same verdict each time. A study of 3.4 million applicants and 4 million applications across 156 employers in 11 sectors gives this failure mode a name: algorithmic monoculture in hiring. The unsettling part runs deeper than algorithms rejecting people: they reject the same people in unison, and every single employer stays blind to it from inside its own funnel.

This is the part of AI-in-hiring that the governance conversation has been missing. We have written about HR becoming a governance function, and about the McKinsey numbers showing how fast AI is moving into people operations. That framing treats each company’s hiring AI as its own risk to manage. The study breaks that assumption. When over 90% of U.S. employers rely on hiring algorithms to screen applicants, and a small number of vendors supply most of them, your hiring-AI exposure is correlated with your competitors’. The system you run is shared, even when it looks like your own.

What Correlated Rejection Actually Does

Picture the labor market as a set of doors. Under independent decisions, a qualified candidate rejected at one door still has real odds at the next, because each employer weighs the application differently. That diversity of judgment is what makes the market function. It is the statistical reason a good candidate eventually lands.

Now replace the independent judgments with one judgment, copied. The study models exactly this, and the result is stark. The number of applications a candidate must submit to land a single offer rises from roughly 10 under independent decisions to 25 under monoculture, a 150% increase. The authors validated this against real audit data from Kline et al. (2022), so the number holds up outside the simulation. The same candidate, the same qualifications, two and a half times the effort, purely because the doors stopped deciding independently and started deciding together.

For the candidate, this reads as bad luck repeated. For the labor market, it is a structural inefficiency: qualified people sitting unhired because the system has no diversity of opinion left to give them a second look. And for the employer, the part that should command attention, it means your rejection of a candidate is no longer your decision alone. It becomes a coordinated outcome beyond your intent and outside your observation.

The Disparate-Impact Surface

Correlated rejection would be a market problem even if it were race-neutral. The study shows it carries measurable racial bias.

When the same screening logic is applied at scale, its biases are applied at scale too. The study found that 25.87% of applications from Black applicants and 14.74% from Asian applicants were routed to positions exhibiting adverse impact. These are the systemic signature of a few algorithms making correlated decisions across a market, rather than isolated vendor bugs. A single biased model, deployed by one employer, harms the people who apply there. The same model, deployed across an entire sector, encodes that bias into the structure of who gets hired anywhere.

This is where the legal exposure lives. U.S. employment law has a long-settled doctrine of disparate impact: a hiring practice that is neutral on its face but produces discriminatory outcomes can be unlawful regardless of intent. “Our vendor’s algorithm did it” fails as a defense. The employer is responsible for the outcomes of the tools it deploys. With HireVue’s algorithms alone used by more than 60% of the Fortune 100, the question for a general counsel is no longer hypothetical. If your screener produces adverse impact, and you lack proof that you tested for it, you own that result in front of a regulator.

Why It Stays Invisible From Inside

The defining feature of monoculture risk is its invisibility to any single participant. Your funnel looks healthy. You receive applications, your screener processes them, you make hires. Every metric you can measure stays inside your own four walls.

What escapes your measurement is the candidate’s experience across all the doors. Hidden from you is the fact that the strong applicant you rejected was also rejected by your three closest competitors running the same vendor. Hidden too is the fact that your screener and theirs share a training lineage that makes their errors your errors. The correlation is real, it is consequential, and it is structurally outside your line of sight. This is why monoculture is a governance problem that reaches past mere model quality. Any tuning of your own algorithm leaves intact the fact that it moves in lockstep with everyone else’s.

Vendor concentration is the mechanism. When a handful of providers supply screening to most of a sector, market-wide correlation becomes the default state, well beyond a risk that might emerge. Diversity of judgment has to be deliberately preserved, and right now almost nobody is preserving it.

What This Changes for How You Govern Hiring AI

Two shifts follow, and both run against current procurement habits.

First, your vendor’s market share is now a risk input. The instinct in enterprise buying is to choose the dominant vendor because everyone uses it, so it must be safe. In hiring AI, that logic inverts. The more concentrated your vendor, the more correlated your decisions are with the rest of the market, and the more your rejections compound into systemic, legally exposed outcomes that sit beyond your observation. Popularity has become the leading risk factor in the choice.

Second, “did we test our model for bias” is the wrong scope. The right question is whether you can demonstrate independent judgment in your hiring decisions at all. If your screener, your competitor’s screener, and a third of your sector all descend from the same vendor model, an internal bias audit of your instance misses the systemic effect entirely. You need to know, beyond whether your tool is fair, whether it is making decisions that are genuinely yours.

Do This Now

Find out what your screening vendor’s market share is in your sector, and treat a high number as a flag. Then run one diagnostic on your hiring pipeline: of the candidates you rejected through automated screening in the last quarter, do you have a documented test for adverse impact across protected groups, and could you produce it for a regulator tomorrow?

If the answer is no, you have a disparate-impact exposure you currently cannot defend, and it is amplified by every competitor running the same tool. The fix demands more than a better model from the same vendor: it takes a deliberate decision about where human judgment re-enters the funnel, so that your rejections are decisions you made rather than outcomes a shared algorithm produced on your behalf. Independence has to be designed back in. Monoculture is the default, and the default is now a liability.

---

This analysis synthesizes Algorithmic Monocultures in Hiring (Algorithmic Monocultures in Hiring, June 2026).

Victorino Group helps enterprises govern AI in hiring and other high-stakes functions before correlated risk becomes legal liability. Let’s talk.