Algorithmic Monocultures in Hiring: 90% of US Employers Share One Vendor
Over 90% of U.S. employers now rely on hiring algorithms to screen job applicants. Most of them buy from just a handful of vendors. A new study, the largest of its kind, analyzes data from 3.4 million real applicants submitting 4 million applications to 156 employers across 11 market sectors. Every application was processed by algorithms from a single vendor. The findings are stark: algorithmic monocultures create systemic rejections and racial disparities at scale.
Key Findings
1. Large-scale adverse impact for Asians and Blacks. This is the first empirical demonstration of adverse impact in deployed algorithmic hiring. 25.87% of applications from Black applicants and 14.74% from Asian applicants were directed to positions where they were adversely impacted under Title VII of the U.S. Civil Rights Act.
2. Adverse impact hidden in aggregate. Prior studies found minimal adverse impact because they looked at all data together. By analyzing each position separately, as Title VII requires, the researchers uncovered disparities that were previously washed out.
3. Systemic rejections exceed random chance. For applicants submitting 4 applications, 10% were rejected everywhere. This rate is significantly higher than the baseline expected if decisions were independent (χ² = 18,481, p < 0.001). The baseline model accurately predicted systemic rejection rates in non-algorithmic hiring data (Kline et al., 2022), confirming that the excess homogeneity is unique to centralized algorithmic assessment.
4. Data access blocks independent research. This study is the first and only independent empirical research on deployed hiring algorithms at scale. The authors argue that policy intervention is needed to enable scientific inquiry and accountability.
How Algorithmic Monocultures Form
Many employers buy hiring algorithms from the same third-party vendors. Over 60% of the Fortune 100 use HireVue's algorithms. When a single vendor's algorithms mediate hiring at multiple employers, they form an algorithmic monoculture. Figure 1 in the paper illustrates the pipeline: job applications are assessed by AI to be recommended or not; if not recommended, the applicant is likely rejected without human review.
Revealing Adverse Impact
Title VII governs discrimination in hiring. Prior studies found limited adverse impact in aggregate. But by studying positions separately, the researchers found that Black applicants are most likely to be adversely impacted: 30% of Black applicants applied to at least one position with adverse impact against them. Asian applicants experienced the largest total shortfall: if Asians were selected at the same rate as the most selected racial group for each position, 29,000 additional Asian applications would have been recommended.
Identifying Systemic Rejection
When applicants apply to multiple positions, they can receive the same outcomes. Algorithmic monoculture makes this more likely. Of applicants that submitted 4 applications, 10% were systemically rejected. The observed rate significantly exceeds the baseline of independent decisions. The baseline was validated using data from Kline et al. (2022), which sent 83,000 applications to 108 Fortune 500 firms. In that non-algorithmic data, the baseline accurately predicted observed rates (χ² = 20.05, p = 0.69). So the excess homogeneity is distinctive of centralized algorithmic assessment.
Simulating Counterfactual Outcomes
What if applicants applied more broadly? When applicants apply everywhere, at least one model recommends them. But under realistic behavior, applicants need to submit 25 applications to ensure at least one recommendation with 99.9% probability — compared to just 10 applications for independent decisions.
Policy Recommendations
Hiring AI is governed by employment discrimination law, general AI regulation, and specific rules. In the U.S., Title VII already evaluates adverse impact at the level of specific jobs. The EU AI Act treats hiring systems as high-risk. New York City Local Law 144 established an early framework, but current guidance doesn't address position-level effects.
The authors recommend:
- Measure adverse impact per position.
- Strengthen market surveillance to quantify homogeneous outcomes.
- Monitor algorithmic monoculture in the hiring supply chain.
- Expand researcher access to data from major hiring platforms.
Technical Details
The study uses data from a single vendor's algorithms processing 4 million applications. The researchers applied the four-fifths rule from Title VII to each position. They computed the systemic rejection rate as the observed proportion of applicants rejected from all positions applied to, compared to a baseline of independent Bernoulli trials with the same per-position rejection probabilities. The chi-squared test showed significant deviation.
What Developers Should Know
If you're building or auditing hiring systems, this paper is essential reading. It demonstrates that aggregate metrics can hide discrimination. Always evaluate per-position adverse impact. The code and data are available at algorithmichiring.github.io.
Why It Matters
Hiring algorithms mediate high-stakes decisions for millions of people. This study shows that algorithmic monocultures can create systemic rejections and racial disparities at scale. Developers building such systems must understand these risks and design for fairness, not just accuracy.
Editor's Take
I've consulted for a few HR tech startups, and this paper confirms what I've long suspected: the industry's reliance on a handful of vendors is a systemic risk. When I see 60% of Fortune 100 companies using HireVue, I worry about the lack of diversity in algorithmic approaches. The finding that aggregate analysis hides adverse impact is a wake-up call. I think regulators should mandate per-position audits, and developers should push back against black-box vendor algorithms. We need more transparency and independent research, not just vendor claims of fairness.
