Algorithmic Monocultures in Hiring
Using pymetrics data, reveals algorithmic monoculture causes racial disparities and systemic rejection in hiring.
Key Findings
Methodology
This study analyzes a novel dataset from the pymetrics platform containing 3.37 million applicants and 4.19 million applications screened by machine learning models from a single vendor. The models use gameplay features from online assessments to generate binary recommendations. The authors apply the U.S. EEOC four-fifths rule to assess adverse racial impact at the granular position level rather than aggregated data. Leveraging the deterministic replicability of the algorithms, they simulate counterfactual outcomes for applicants applying to all positions, enabling quantification of systemic rejection and the effects of algorithmic monoculture.
Key Results
- 14.74% of applications by Asian applicants and 25.87% by Black applicants were submitted to positions exhibiting adverse impact against these groups, indicating significant racial disparities.
- 4% of applicants who applied to 10 positions were rejected by all, a rate substantially higher than expected under independent decision baselines, demonstrating outcome homogenization due to algorithmic monoculture.
- Counterfactual simulations show that if applicants applied widely, each would receive at least one recommendation; however, in practice, some remain systemically rejected, requiring about 25 applications to reduce systemic rejection below 0.1%.
Significance
This work provides the first large-scale empirical evidence of how algorithmic monoculture in hiring—where many employers rely on the same vendor’s algorithms—produces systemic racial disparities and homogeneous rejection outcomes. It challenges prior aggregate analyses that mask position-level discrimination and highlights risks inherent in vendor concentration. The findings have profound implications for both academia and industry, emphasizing the need for transparency, independent auditing, and policy interventions to ensure fairness in algorithmic hiring.
Technical Contribution
The paper introduces a novel methodological framework leveraging the deterministic nature of hiring algorithms to simulate counterfactual applicant outcomes across all positions, overcoming traditional social science limitations in cross-employer tracking. It provides a fine-grained position-level adverse impact analysis and quantifies systemic rejection rates, contrasting observed outcomes with independent decision baselines. This approach enables precise measurement of algorithmic monoculture effects and offers actionable insights for mitigating systemic bias.
Novelty
This is the first study to use real-world, multi-employer data from a single algorithm vendor at this scale to empirically demonstrate algorithmic monoculture’s role in systemic racial bias and rejection homogenization. Unlike prior work focused on resumes or single employers, it uniquely exploits game-based assessment data and algorithmic determinism to reveal new dimensions of hiring algorithm fairness.
Limitations
- Over 62% of applicants did not self-report race, potentially underestimating racial disparities and limiting proxy discrimination analysis.
- Data is restricted to pymetrics and its clients, so findings may not generalize to other vendors or hiring contexts.
- Binary thresholding of recommendation scores ignores nuanced recommendation levels used by some clients, which may affect the generalizability of rejection rate estimates.
Future Work
Future research should extend to multiple algorithm vendors and platforms, incorporate richer applicant features and dynamic behaviors, and explore the formation and mitigation of algorithmic monoculture. Policy efforts should focus on enhancing transparency and enabling independent audits to foster equitable hiring practices.
AI Executive Summary
The widespread adoption of AI-driven hiring algorithms has transformed recruitment, with over 90% of U.S. employers relying on automated screening. However, many employers procure these algorithms from a handful of vendors, raising concerns about algorithmic monoculture—where identical or similar algorithms mediate hiring decisions across multiple employers. This study leverages a unique dataset from pymetrics, encompassing over 3.3 million applicants and 4.1 million applications across 156 employers and 1,746 positions, to investigate the implications of such monoculture on fairness and systemic rejection.
The authors apply the U.S. Equal Employment Opportunity Commission’s (EEOC) four-fifths rule to assess adverse impact at the position level, revealing that 14.74% of applications by Asian applicants and 25.87% by Black applicants were submitted to positions exhibiting adverse impact against these groups. Notably, 10.62% of positions demonstrated adverse impact against Black applicants, affecting nearly a third of Black applicants. These disparities were obscured in prior aggregate analyses, underscoring the importance of granular evaluation.
Beyond racial disparities, the study uncovers systemic rejection patterns: 4% of applicants who applied to 10 positions were rejected by all, a rate significantly exceeding what would be expected if decisions were independent. This outcome homogenization stems from the deterministic nature of the algorithms and their reuse across positions, exemplifying the risks of algorithmic monoculture. To further understand this, the authors simulate counterfactual outcomes assuming applicants applied to all positions, finding that theoretically every applicant would receive at least one recommendation. However, in practice, some remain systemically rejected, necessitating broader application strategies to mitigate this risk.
This research breaks new ground by combining large-scale, multi-employer data with algorithmic determinism to quantify the systemic effects of hiring algorithms. It challenges existing fairness assessments that rely on aggregated data and single-employer studies, highlighting the need for transparency and independent research. The findings have significant implications for vendors, employers, and policymakers aiming to foster equitable hiring in an increasingly automated landscape.
While comprehensive, the study acknowledges limitations including incomplete race reporting and vendor-specific data scope. Future work should broaden the scope to other platforms, incorporate richer applicant data, and explore policy frameworks to address algorithmic monoculture. Overall, this work provides critical insights into the complex interplay between algorithm design, vendor concentration, and labor market fairness.
Deep Analysis
Background
The use of machine learning algorithms in hiring has become ubiquitous, with over 90% of U.S. employers employing automated screening or ranking systems to filter job applicants. These algorithms aim to improve efficiency and objectivity in recruitment but have raised concerns about fairness and bias. Prior research has documented discriminatory patterns in resume-based screening, often linked to proxies for sensitive attributes such as race or gender. However, these studies typically focus on single employers or rely on synthetic data, limiting their ability to capture systemic effects across the labor market. Concurrently, the market for hiring algorithms is highly concentrated, with a few vendors like pymetrics and HireVue serving a large share of employers. This concentration creates an algorithmic monoculture, where identical or similar models influence hiring decisions across multiple organizations. The implications of such monoculture for systemic bias and applicant outcomes remain underexplored due to data access challenges and methodological constraints.
Core Problem
The central problem addressed is whether algorithmic monoculture leads to systemic rejection of the same individuals across multiple job applications, particularly affecting racial minorities. Traditional fairness assessments aggregate data across positions and employers, potentially masking position-specific adverse impacts. Moreover, the deterministic nature of hiring algorithms reused across positions may produce homogeneous outcomes, reducing applicants’ chances of success. Understanding these dynamics is crucial because systemic rejection can exacerbate unemployment, discourage job seeking, and distort labor market efficiency. The challenge lies in obtaining and analyzing large-scale, cross-employer data and developing methodologies to simulate counterfactual outcomes to disentangle the effects of application breadth and algorithmic design on fairness.
Innovation
This work introduces several key innovations:
- �� It leverages a unique, large-scale dataset from pymetrics covering over 3 million applicants and 4 million applications across 156 employers and 1,746 positions, enabling unprecedented cross-employer analysis.
- �� It applies the EEOC four-fifths rule at the granular position level rather than aggregating data, revealing hidden racial disparities.
- �� It exploits the deterministic replicability of the hiring algorithms to simulate counterfactual applicant outcomes across all positions, overcoming traditional social science limitations in tracking applicants across employers.
- �� It quantifies systemic rejection rates and contrasts observed outcomes with a baseline assuming independent decisions, thereby isolating the effects of algorithmic monoculture.
These contributions collectively advance the understanding of how vendor concentration and algorithm design shape labor market fairness.
Methodology
- �� Data Collection: Obtained from pymetrics platform, spanning Dec 2018 to Dec 2022, including 3,372,132 applicants, 4,197,168 applications, 156 employers, and 1,746 positions.
- �� Algorithmic Screening: Each application is evaluated by a binary classifier trained per position, using gameplay features from 12 standardized online games. Positive training examples are profiles of at least 50 current employees; negatives are random profiles.
- �� Recommendation Output: The classifier outputs a score in [0,1], thresholded at 0.5 to produce binary recommendations (‘recommend’ or ‘do not recommend’). ‘Do not recommend’ implies likely rejection without human review.
- �� Race Data: 40.2% of applicants self-report race, categorized mainly as Asian, Black, White, Hispanic.
- �� Adverse Impact Assessment: Calculated selection rates per racial group per position, computed impact ratios relative to the most selected group, applying the EEOC four-fifths rule and Benjamini-Hochberg correction for multiple testing.
- �� Systemic Rejection Analysis: Measured the proportion of applicants rejected by all positions they applied to, compared against a baseline assuming independent decisions.
- �� Counterfactual Simulation: Leveraged deterministic algorithm outputs to simulate recommendations if applicants applied to all positions, estimating the effect of application breadth on systemic rejection.
Experiments
Experiments utilized the large-scale pymetrics dataset covering diverse industries including finance, manufacturing, and technology. The average position received 2,404 applications. The authors analyzed racial disparities using the EEOC four-fifths rule at the position level, identifying positions with statistically significant adverse impact. Systemic rejection rates were computed for applicants applying to multiple positions and compared to a baseline model assuming independent decisions. Counterfactual simulations generated recommendation outcomes for hypothetical applications to all positions, evaluating how application breadth influences systemic rejection. The study also examined model sharing across employers to assess total algorithmic monoculture effects. Statistical significance was controlled via Benjamini-Hochberg correction.
Results
Key findings include: Asian and Black applicants face significant adverse impact at the position level, with 10.62% of positions adversely impacting Black applicants and over 30% of Black applicants applying to at least one such position. Approximately 25.87% of Black applications were to positions with adverse impact. Systemic rejection was observed in 4% of applicants applying to 10 positions, exceeding independent decision baselines. Counterfactual simulations showed that if applicants applied to all positions, each would receive at least one recommendation, but in practice, some remain systemically rejected, requiring about 25 applications to reduce systemic rejection below 0.1%. These results highlight the pervasive and systemic nature of algorithmic monoculture effects in hiring.
Applications
The findings have immediate applications for algorithm vendors, employers, and policymakers. Vendors can use the proposed position-level adverse impact and counterfactual simulation methods to audit and improve their models. Employers should complement algorithmic screening with human oversight to mitigate systemic exclusion. Policymakers can leverage these insights to craft regulations enhancing transparency, fairness, and independent auditing of hiring algorithms. Academically, the study provides a novel framework and dataset for cross-employer algorithmic fairness research, fostering interdisciplinary collaboration and innovation in equitable AI deployment.
Limitations & Outlook
The study faces several limitations: First, 62.35% of applicants did not report race, potentially biasing disparity estimates and limiting analysis of proxy discrimination. Second, data is confined to pymetrics and its clients, restricting generalizability to other vendors or hiring contexts. Third, the binary thresholding of recommendations overlooks nuanced multi-tier recommendation schemes used by some clients, which may affect rejection rate interpretations. Additionally, lack of access to applicant names and other sensitive features limits deeper investigation into indirect discrimination mechanisms.
Abstract
Many employers screen job applicants with algorithms built by the same few algorithm vendors. We hypothesize that algorithmic monoculture leads to the same individuals and members of the same racial groups facing rejection. We acquire and analyze a novel dataset of 3 million applicants submitting 4 million applications where all the applications are screened by algorithms built by the same vendor. We find clear racial disparities in applicant outcomes. Of all applications submitted by Asian and Black applicants, 14.74% and 25.87% are submitted to positions that adversely impact Asian and Black applicants, respectively, according to U.S. employment discrimination standards. Individuals also receive homogeneous outcomes: 4% of all applicants who apply to 10 positions are recommended for rejection from all positions, a rate higher than expected by chance. To better understand this homogeneity, we leverage the deterministic replicability of hiring algorithms to generate the outcomes applicants would have received if they applied to all positions. We show that applicants would need to apply widely in order to ensure their applications are considered by a human
References (20)
Systemic Discrimination Among Large U.S. Employers
Patrick M. Kline, Evan K. Rose, Christopher R. Walters
Algorithmic monoculture and social welfare
J. Kleinberg, Manish Raghavan
Duration Dependence and Labor Market Conditions: Evidence from a Field Experiment*
Kory Kroft, Fabian Lange, Matthew J. Notowidigdo
Advancing Science- and Evidence-based AI Policy
Rishi Bommasani, Sanjeev Arora, Jennifer T. Chayes et al.
Private Government: How Employers Rule Our Lives (and Why We Don’t Talk About It)
G. Donleavy
Mismatch Unemployment and the Geography of Job Search
Ioana Marinescu, Roland Rathelot
Introducing contextual transparency for automated decision systems
Mona Sloane, Ian Solano-Kamaiko, Jun Yuan et al.
When the Machine Meets the Expert: An Ethnography of Developing AI for Hiring
Elmira van den Broek, Anastasia V. Sergeeva, M. Huysman
AI for hiring in context: a perspective on overcoming the unique challenges of employment research to mitigate disparate impact
Sara Kassir, Lewis Baker, J. Dolphin et al.
Incentivized Resume Rating: Eliciting Employer Preferences without Deception
Judd B. Kessler, Corinne Low, Colin D. Sullivan
Ecosystem-level Analysis of Deployed Machine Learning Reveals Homogeneous Outcomes
Connor Toups, Rishi Bommasani, Kathleen A. Creel et al.
Confronting Power and Corporate Capture at the FAccT Conference
Meg Young, Michael A. Katell, P. Krafft
Uniform Guidelines on Employee Selection Procedures
Teva J. Scheer
The Allocation of Talent and U.S. Economic Growth
Chang-tai Hsieh, Erik Hurst, Charles I. Jones et al.
An analysis of long-term unemployment
D. Rothstein
Gender-targeted job ads in the recruitment process: Facts from a Chinese job board
P. Kuhn, Kailing Shen, Shuo Zhang
Does Job Testing Harm Minority Workers? Evidence from Retail Establishments
David H. Autor, D. Scarborough
A Silicon Valley love triangle: Hiring algorithms, pseudo-science, and the quest for auditability
Mona Sloane, E. Moss, Rumman Chowdhury
The hard proxy problem: proxies aren’t intentional; they’re intentional
Gabbrielle M. Johnson
Inequality, Unemployment and Contemporary Europe.
A. Sen