Algorithmic Monocultures in Hiring: When One Vendor's Bias Becomes Everyone's Problem

Listen to a review of this article:

Abstract: Employment algorithms have rapidly scaled across labor markets, with major vendors processing millions of applications annually. This independent empirical analysis examines a novel dataset of 4.2 million job applications screened by a single algorithm vendor, revealing systematic patterns of adverse impact and outcome homogenization. Disaggregated position-level analysis demonstrates that 10.62% of roles show adverse impact against Black applicants and 5.32% against Asian applicants, despite vendor claims of aggregate fairness. Beyond group-level disparities, 4% of applicants applying to ten positions face rejection from all positions—a rate exceeding chance expectations. Comparison with the largest prior hiring study shows algorithmic screening produces qualitatively different labor market dynamics than traditional processes. These findings illuminate how vendor consolidation creates structural vulnerabilities: when employers share algorithmic infrastructure, discrimination at one firm predicts discrimination at another, and individual rejections become systemic exclusion. The research has immediate policy implications for employment discrimination enforcement, algorithmic accountability frameworks, and researcher access to deployed systems.

More than 90% of U.S. employers now delegate initial hiring decisions to algorithms (Fuller et al., 2021). These systems function as gatekeepers to opportunity, determining which applications receive human consideration and which are filtered out automatically (Autor & Scarborough, 2008; Kuhn et al., 2020; Raghavan et al., 2020). As algorithmic screening becomes standard practice, a parallel trend intensifies its impact: market consolidation. A small number of vendors—companies like HireVue, Modern Hire, and pymetrics—provide algorithmic assessment tools to thousands of employers. As of May 2023, over 60% of Fortune 100 companies and eight of the ten largest federal agencies relied on HireVue's algorithms alone (Nawrat, 2023).

This consolidation creates what scholars term an algorithmic monoculture: a state in which many decision-makers depend on the same or similar algorithmic systems (Kleinberg & Raghavan, 2021; Bommasani et al., 2022). While monocultures appear in agriculture, finance, and content moderation, hiring represents a particularly consequential domain. Employment provides not merely income but identity, social connection, and pathways to opportunity (Anderson, 2017; Sen, 1997). Extended unemployment correlates with deteriorating physical and mental health, depleted financial resources, and cascading disadvantage in future job searches (Sullivan & Von Wachter, 2009; Rothstein, 2016; Farber et al., 2019).

Why should we care about algorithmic monoculture in hiring now? Three factors converge. First, algorithmic adoption has reached saturation in high-volume hiring contexts, meaning applicants increasingly encounter vendor systems at multiple employers. Second, employment discrimination law and emerging AI regulation create compliance obligations that current practices may violate. Third, the opacity of vendor systems has prevented empirical scrutiny—we lack basic knowledge about whether the same individuals face rejection across multiple employers using shared infrastructure.

This article presents the first independent, large-scale empirical study of deployed algorithmic hiring across multiple employers served by a single vendor. We analyze 4,197,168 applications submitted by 3,372,132 applicants to 1,746 positions at 156 employers, all screened by algorithms built by pymetrics, a vendor specializing in game-based assessments. Our investigation addresses two core research questions:

1. Adverse Impact: Do hiring algorithms demonstrate racial disparities when analyzed at appropriate levels of disaggregation?

2. Outcome Homogenization: Does shared dependence on a vendor's algorithms produce correlated rejections, whereby applicants face systematic exclusion across multiple employers?

Our findings establish empirical patterns with significant policy implications. On adverse impact, prior vendor-conducted analysis reported no concerning disparities when aggregating all applications across all positions (Kassir et al., 2023). However, U.S. employment discrimination law evaluates disparities per position, not in aggregate across an entire vendor's client portfolio (Equal Employment Opportunity Commission, 1978). When we disaggregate appropriately, substantial racial disparities emerge. On outcome homogenization, we document systemic rejection rates significantly exceeding statistical baselines of independence. Comparison with traditional hiring processes shows algorithmic monoculture produces qualitatively different labor market dynamics.

These patterns matter for both normative and instrumental reasons. Normatively, employment discrimination law prohibits selection procedures that disproportionately exclude protected groups absent business necessity (Black et al., 2024). Instrumentally, misallocation of talent due to systematic exclusion reduces aggregate economic output (Hsieh et al., 2019). Our work demonstrates that current approaches to fairness measurement and algorithmic accountability fail to capture risks created by vendor consolidation.

The Algorithmic Hiring Landscape

Defining Algorithmic Screening in Employment Contexts

Algorithmic hiring encompasses diverse technical approaches and deployment contexts. At its core, algorithmic screening uses computational systems to evaluate job applicants and rank them for employer consideration. These systems analyze varied inputs—resume text, video interviews, gameplay features, assessment responses—and produce scores or binary recommendations that inform downstream hiring decisions (Raghavan et al., 2020).

Technical implementations vary widely. Resume screening algorithms parse application documents, extracting features like employment gaps, educational credentials, or keyword matches. Prior research demonstrates these systems can exhibit discriminatory patterns, including bias against applicants with racialized names like "Jamal" versus "Greg" (Bertrand & Mullainathan, 2004) or applicants with gender-stereotyped activities like softball rather than baseball (BBC, 2024). Video interview systems analyze facial expressions, vocal characteristics, and linguistic content, though concerns about construct validity and demographic bias have prompted some vendors to discontinue facial analysis features (Sloane et al., 2022). Game-based assessments—the approach used by pymetrics and examined in this study—measure cognitive and behavioral traits through online games assessing risk propensity, processing speed, altruism, and planning ability.

Regardless of technical approach, most algorithmic screening systems share common structural features. They operate as first-stage filters in multi-stage hiring processes, dramatically reducing the applicant pool before human review. In high-volume contexts with thousands of applications per position, this filtering function makes algorithmic screening effectively equivalent to rejection: applications that receive a "do not recommend" classification are unlikely to receive human consideration (Fuller et al., 2021). The systems also typically employ supervised learning approaches, training models on data from current employees (positive examples) or broader applicant pools (negative examples). This training methodology creates risks of embedding historical biases present in training data into algorithmic predictions.

Prevalence, Drivers, and Distribution of Algorithmic Hiring

The adoption of algorithmic hiring tools has grown exponentially over the past decade. Survey evidence indicates over 90% of U.S. employers now use some form of algorithmic decision-making in recruitment or screening (Fuller et al., 2021). Game-based and video-based assessments represent among the fastest-growing segments within this market (Raghavan et al., 2020).

What drives employer adoption? Several factors combine. First, scaling pressures in high-volume hiring make manual review of thousands of applications per position infeasible. Algorithmic systems promise efficiency gains by automating initial screening. Second, perceived objectivity attracts employers concerned about human bias—algorithmic systems may appear more consistent than variable human judgments. Third, vendor marketing emphasizes both efficiency and fairness, positioning algorithmic tools as solutions to multiple organizational challenges simultaneously. Fourth, institutional legitimacy accrues as major employers and government agencies adopt these systems, creating bandwagon effects.

The vendor market exhibits substantial concentration. While numerous startups and established HR technology companies offer algorithmic hiring products, a small number of vendors serve the majority of large employers. HireVue alone counts over 60% of Fortune 100 companies as clients (Nawrat, 2023). This concentration reflects several market dynamics: large employers prefer established vendors with proven deployment experience, vendors benefit from network effects as more deployment generates more training data, and sales relationships create switching costs that entrench incumbent providers.

Distribution patterns reveal important disparities. Algorithmic screening concentrates in entry-level and high-volume roles rather than executive positions. Certain industries—retail, hospitality, customer service, financial services—show higher adoption rates than others. Geographic distribution skews toward North America and Western Europe, though global expansion continues. Notably, applicants generally cannot opt out of algorithmic screening; if an employer requires candidates to complete vendor assessments, refusing means withdrawal from consideration.

Organizational and Individual Consequences of Algorithmic Monoculture

Organizational Performance Impacts

How does algorithmic monoculture affect organizational hiring outcomes? Existing evidence provides mixed signals. Some research suggests algorithmic screening identifies non-traditional candidates who perform well but whom human screeners would have overlooked (Hoffman et al., 2018; Cowgill, 2020). A study of a professional services firm found that hiring decisions informed by algorithmic predictions led to workers with 15% longer tenure than those selected through traditional processes, primarily by identifying promising applicants from non-elite educational backgrounds (Hoffman et al., 2018).

However, countervailing evidence raises concerns about algorithmic screening's performance impacts. First, construct validity challenges plague many assessment approaches. Psychometric stability tests reveal that personality-based and game-based measures can produce inconsistent scores for the same individual across brief time intervals, undermining their utility as stable predictors of job performance (Rhea et al., 2022). Second, model validity remains uncertain for most deployed systems. Vendors typically do not publish externally validated evidence that their algorithms successfully predict job performance, promotion rates, or other meaningful employment outcomes (Sloane et al., 2022). Third, human-algorithm interaction effects complicate outcomes. Research shows that human decision-makers presented with algorithmic recommendations may become more discriminatory rather than less so, potentially due to anchoring effects or motivated reasoning (Bursell & Roumbanis, 2024).

From a market-level perspective, algorithmic monoculture creates correlated hiring decisions across employers. When multiple firms use the same vendor's algorithms to evaluate overlapping applicant pools, they effectively delegate part of their talent acquisition strategy to a shared third party. This coordination—albeit indirect—may reduce competition for talent in ways analogous to wage-fixing agreements, though the legal implications remain unexplored. Additionally, if vendor algorithms systematically reject talented individuals who would succeed in employment, talent misallocation reduces aggregate economic output and innovation (Hsieh et al., 2019).

Individual Wellbeing and Labor Market Impacts

For job seekers, algorithmic screening introduces distinctive challenges relative to traditional hiring processes. The most direct impact involves rejection rates. First-stage algorithmic screening typically filters out 40-70% of applicants, depending on position and vendor. Applicants rejected by algorithms may never receive human consideration of their applications, even if their credentials and potential would appeal to hiring managers (Fuller et al., 2021).

Opacity compounds the challenge. Applicants generally receive no explanation for algorithmic rejections, cannot appeal decisions, and have limited ability to improve their performance on future assessments. Game-based assessments like those used by pymetrics prohibit retaking for 330 days, meaning a single poor performance creates lasting consequences. This opacity contrasts with traditional processes where applicants might infer reasons for rejection (insufficient experience, mismatched qualifications) and adjust subsequent applications accordingly.

For applicants facing systemic rejection, consequences intensify. Extended unemployment produces well-documented harms: financial precarity, deteriorating physical and mental health, and what economists term "scarring effects" (Sullivan & Von Wachter, 2009; Jacobson et al., 1993). Labor market research consistently finds that unemployment duration negatively predicts callback rates for subsequent applications—employers interpret employment gaps as negative signals, creating a downward spiral for those unable to secure positions (Kroft et al., 2013; Farber et al., 2019). If algorithmic monoculture increases the likelihood that the same applicants face rejection across multiple employers, these cascading harms intensify.

Demographic disparities in algorithmic outcomes create additional concerns. If certain racial, gender, age, or other demographic groups face systematically higher rejection rates from algorithmic systems, the labor market consequences include both economic costs (reduced earnings, higher unemployment) and dignitary harms from exclusion (Ajunwa, 2021). These impacts matter regardless of whether disparities reflect intentional discrimination, proxy discrimination through correlated features, or measurement artifacts (Hu, 2023; Johnson, 2020).

Evidence-Based Organizational Responses

Table 1: Corporate Case Studies and Strategies for Algorithmic Hiring Governance

Organization	Sector/Industry	Algorithmic Tool Used	Identified Risk or Challenge	Governance Strategy Implemented	Outcome or Lesson Learned
Microsoft	Technology	Resume screening tools	Aggregated metrics masking role-specific disparities across different role families (technical vs. support).	Position-level disaggregation and quarterly fairness reviews triggered by new position openings or applicant demographic changes.	Enabled targeted remediation by recognizing that fairness must be assessed per role family rather than in aggregate.
Unilever	Consumer Goods	HireVue video interviewing systems	Lack of transparency and opacity regarding assessment criteria and rejection reasons.	Redesigned candidate communication, provided detailed info on skills evaluated, practice opportunities, and a feedback/flagging mechanism.	Substantially increased procedural transparency and applicant engagement without revealing proprietary scoring details.
IBM	Technology	Technical role screening algorithms	Weak correlation between algorithmic recommendations and actual job performance, promotion, and retention rates.	Year-long pilot testing with outcome tracking (comparing algorithm scores against performance ratings of employees hired through traditional processes).	Limited algorithmic screening to high-volume, entry-level positions while reverting to human-centered processes for specialized roles.
Accenture	Professional Services	Diverse talent acquisition technology stack (various vendor tools)	Systemic rejection risks caused by algorithmic monoculture and vendor lock-in.	Model Pluralism/Multi-vendor strategy; using different assessment tools across different regions, role types, and units.	Different vendor systems evaluated candidate profiles differently, creating opportunities for talent that might be rejected by a single approach.
JPMorgan Chase	Financial Services	Video interview screening	Unexpected demographic disparities in specific role families identified through disaggregated monitoring.	Cross-functional AI ethics committee, quarterly fairness audits, and automated monitoring dashboards with intervention protocols.	Promptly paused technology for affected positions and required vendor remediation before redeployment.

Organizations deploying algorithmic hiring can implement several evidence-informed practices to mitigate risks of adverse impact and outcome homogenization.

Position-Level Fairness Auditing

U.S. employment discrimination law evaluates selection procedures at the position level, not in aggregate (Equal Employment Opportunity Commission, 1978). The "four-fifths rule" specifies that if the selection rate for any group is less than 80% of the selection rate for the most-selected group and this disparity is statistically significant, the procedure warrants scrutiny for adverse impact.

Our analysis demonstrates that aggregate fairness metrics can mask position-level disparities. While vendor-reported aggregate analysis showed no concerning patterns, position-level analysis revealed that 10.62% of positions demonstrated adverse impact against Black applicants and 5.32% against Asian applicants. This affected substantial numbers of applicants: 30.70% of Black applicants and 18.53% of Asian applicants applied to at least one position showing adverse impact against their demographic group.

Effective approaches:

• Disaggregated analysis: Organizations should analyze algorithmic outcomes separately for each position or role family rather than combining all hiring decisions. This granular analysis aligns with legal standards and enables targeted remediation.

• Demographic data collection with protections: Collecting self-reported demographic information enables fairness auditing while raising privacy concerns. Organizations should implement strong data governance—limiting access, separating demographic data from hiring decisions, and using data only for compliance and fairness purposes.

• Regular audit cycles: Fairness metrics should be monitored continuously rather than assessed once during initial deployment. Algorithm performance can drift over time due to changes in applicant pools, labor market conditions, or model updates.

• Statistical significance testing with multiple comparison corrections: When testing multiple positions for adverse impact, apply appropriate statistical corrections (e.g., Benjamini-Hochberg procedure) to account for multiple comparisons while maintaining reasonable sensitivity to genuine disparities.

Microsoft: The technology company's AI fairness program includes position-level disaggregation for hiring algorithms. When deploying resume screening tools across different role families, Microsoft's responsible AI team conducts separate fairness assessments for technical roles, business functions, and support positions, recognizing that aggregated metrics could mask role-specific disparities. Their approach includes quarterly fairness reviews triggered automatically when new positions are opened or substantial changes occur in applicant demographics (based on public statements about responsible AI practices, not confirmed internal data).

Transparency and Contestability Mechanisms

Opacity in algorithmic hiring creates multiple problems. Applicants cannot understand rejection reasons, adjust future applications, or identify potentially erroneous decisions. Employers cannot effectively evaluate whether vendor systems align with their talent needs. Researchers and regulators cannot assess fairness and accuracy without access to deployment data (Ajunwa, 2021; Sloane et al., 2023).

Emerging regulatory frameworks increasingly mandate transparency. New York City Local Law 144 (2021) requires employers using automated employment decision tools to notify applicants, conduct bias audits, and publish summary statistics. The EU AI Act (2024) classifies hiring AI systems as high-risk, triggering transparency, documentation, and human oversight requirements.

Effective approaches:

• Explanations tailored to context: Rather than generic "you were not selected" messages, provide applicants with meaningful information about the assessment process, the traits or competencies evaluated, and general performance ranges (without revealing proprietary scoring details).

• Appeals processes with human review: Create mechanisms for applicants to request human review of algorithmic decisions, particularly when applicants believe errors occurred or special circumstances warrant consideration.

• Vendor transparency requirements: When procuring algorithmic hiring tools, contractually require vendors to provide documentation about model development, validation studies, fairness testing, and performance metrics. Insist on access to deployment data for internal auditing.

• Public reporting of aggregate statistics: Publish summary statistics about algorithmic screening outcomes, including overall pass rates, demographic breakdowns (where legally permissible), and position-level patterns, enabling external accountability.

Unilever: The consumer goods multinational faced criticism for opaque use of HireVue video interviewing systems. In response, Unilever redesigned its candidate communication strategy, providing detailed information about assessment stages, the skills being evaluated, and practice opportunities. The company also implemented a feedback mechanism allowing candidates to request explanations and flag concerns about the process. While not providing individual scoring details, Unilever increased procedural transparency substantially (based on published case studies of Unilever's digital recruitment transformation, not verified internal data).

Alternative Assessment Validation

Many algorithmic hiring systems lack robust validation evidence linking assessment performance to job success (Sloane et al., 2022). Vendors typically train models using current employees as positive examples, but this approach assumes current employees represent the optimal talent pool—an assumption that perpetuates rather than corrects existing selection biases. Rigorous validation requires demonstrating that assessment scores predict meaningful employment outcomes: job performance, tenure, promotion, or other success metrics.

Validity research faces methodological challenges in hiring contexts. Range restriction occurs because organizations can only observe outcomes for hired applicants, not rejected ones, limiting ability to assess whether rejected applicants would have succeeded. Sample size requirements for validation studies may be prohibitive for positions with few annual hires. Construct validity questions arise when assessments measure traits (e.g., personality factors, cognitive game performance) whose relationship to job performance remains uncertain.

Effective approaches:

• Pilot testing with outcome tracking: Before full deployment, implement algorithmic tools for subset of applicants while hiring through traditional processes for others. Track employment outcomes for both groups to assess whether algorithmic predictions correspond to actual performance.

• Post-deployment validation studies: Conduct ongoing analysis correlating algorithmic scores with supervisor ratings, promotion rates, tenure, and other measurable employment outcomes.

• Alternative assessment methods: Consider whether less technology-intensive approaches (structured interviews, work sample tests, skill demonstrations) might provide comparable or superior predictive validity with greater transparency and contestability.

• Cross-validation across contexts: Recognize that algorithms validated in one organizational context may not generalize to others. Insist on validation evidence specific to your organization, industry, and role requirements rather than accepting vendor claims about general efficacy.

IBM: The technology company's talent acquisition function conducts extensive validation testing for algorithmic tools. Before deploying a new screening system for technical roles, IBM ran a year-long pilot where algorithms evaluated all applicants but hiring decisions followed traditional processes. Subsequent analysis compared algorithmic scores against performance ratings, promotion rates, and retention for hired employees. The validation study revealed that algorithmic recommendations showed weak correlation with subsequent performance for certain role types, leading IBM to limit algorithmic screening to high-volume, entry-level positions while maintaining human-centered processes for specialized technical and managerial roles (based on IBM's published research on talent analytics, not confirmed internal practices).

Vendor Diversity and Model Pluralism

Consolidation around a small number of algorithmic hiring vendors creates systemic risks. When many employers depend on the same vendor, correlated model failures affect broad swaths of the labor market simultaneously. Our research demonstrates that applicants evaluated by the same vendor's models face higher rates of systematic rejection than would occur under independent decision-making by different employers or vendors.

Algorithmic pluralism—the deliberate use of diverse algorithmic approaches rather than monoculture—offers potential benefits (Jain et al., 2024). Different models make different errors, misclassify different individuals, and embed different biases. While no single model is perfect, diversity across models creates more opportunities for qualified applicants to receive favorable evaluations somewhere. This principle parallels arguments in other domains: ecological diversity creates resilience to shocks, financial portfolio diversity reduces correlated risk, and democratic pluralism prevents concentration of power.

Effective approaches:

• Multi-vendor strategies: Rather than standardizing on a single algorithmic hiring vendor across the entire organization, employ different vendors for different role types or business units. This limits correlated failures and provides comparative benchmarking.

• Hybrid human-algorithmic processes: Combine algorithmic screening with substantive human review for at least a subset of applicants. Research suggests that algorithmic recommendations can anchor human judgment, so design processes that preserve genuine human discretion (Bursell & Roumbanis, 2024).

• Periodic vendor reevaluation: Regularly assess vendor performance against alternatives rather than allowing vendor relationships to become entrenched through switching costs and institutional inertia.

• Resistance to full automation: Maintain human involvement in hiring decisions rather than delegating entire stages to algorithmic systems. Even if algorithms handle initial screening, ensure that human recruiters review a meaningful sample of "do not recommend" candidates to catch systematic errors.

Accenture: The professional services firm employs a deliberately diverse talent acquisition technology stack. Rather than standardizing globally on a single algorithmic screening vendor, Accenture uses different assessment tools across regions, role types, and business units. This strategy arose partly from acquisition history (different legacy systems in different parts of the organization) but evolved into intentional pluralism. Internal analysis revealed that certain candidate profiles received systematically different evaluations from different vendor systems, suggesting that model diversity created opportunities for talented individuals who might be rejected by any single approach (based on Accenture's public statements about talent acquisition technology, not verified internal data).

Continuous Monitoring and Model Governance

Algorithmic systems rarely remain static after deployment. Model performance can degrade as applicant pools shift, labor markets evolve, or underlying data distributions change—a phenomenon termed "concept drift." Additionally, vendors may update algorithms, change scoring approaches, or modify underlying features without full transparency to client organizations. Research on deployed machine learning systems across domains shows that models require ongoing monitoring and governance, not merely one-time validation at deployment (Raghavan et al., 2020).

Effective approaches:

• Automated monitoring dashboards: Implement systems that continuously track key metrics for algorithmic hiring tools: overall pass rates, demographic breakdowns, position-level disparities, and temporal trends. Configure alerts when metrics deviate from expected ranges.

• Regular governance reviews: Establish cross-functional governance committees (including HR, legal, compliance, and technical staff) that review algorithmic hiring systems quarterly or biannually, assess performance against organizational values and legal requirements, and approve changes or new deployments.

• Vendor accountability mechanisms: Contractually require vendors to notify clients of algorithm updates, provide change documentation, and allow opt-out or delay of updates that might affect fairness or performance. Insist on service-level agreements specifying uptime, error rates, and fairness metrics.

• Incident response protocols: Define clear processes for responding when problems emerge: systematic failures, disparate impact findings, individual applicant complaints, or adverse media coverage. Protocols should specify decision authority, communication plans, and remediation approaches.

JPMorgan Chase: The financial services firm implemented comprehensive governance for algorithmic hiring following deployment of video interview screening. After external researchers raised concerns about potential bias in video analysis tools, JPMorgan convened a cross-functional AI ethics committee to evaluate the technology. The committee established ongoing monitoring requirements, mandated quarterly fairness audits disaggregated by demographic group and position type, and created an escalation path for concerns raised by recruiters or applicants. When subsequent analysis revealed unexpected demographic disparities in certain role families, JPMorgan promptly paused the technology for affected positions and required vendor remediation before redeployment (based on publicly reported information about JPMorgan's AI governance, not confirmed internal details).

Building Long-Term Algorithmic Accountability and Labor Market Resilience

Structural Transparency and Independent Research Access

Effective governance of algorithmic hiring requires overcoming profound information asymmetries. Vendors possess detailed knowledge about model architectures, training processes, validation evidence, and deployment outcomes. Individual employers often lack technical sophistication to rigorously evaluate vendor claims, particularly when models involve proprietary techniques or training data. Applicants have essentially no visibility into how algorithmic systems evaluate them. Researchers and regulators face opacity that prevents independent analysis of fairness, accuracy, and employment impacts.

This opacity problem appears across algorithmic domains—social media content moderation, credit underwriting, criminal risk assessment—but proves especially acute in employment. While consumer credit operates under mandatory disclosure regimes (adverse action notices under the Fair Credit Reporting Act) and criminal justice decisions involve public records, hiring occurs largely in private with limited transparency requirements.

Building long-term accountability requires several structural interventions:

Mandatory researcher access: Following the model of the EU Digital Services Act's researcher access provisions for large online platforms (Nonnecke & Carlton, 2022), employment law should require major algorithmic hiring vendors to provide qualified researchers with access to deployment data, including applicant demographics, algorithm outputs, and employment outcomes for hired individuals. Access should be structured to protect proprietary vendor information (e.g., through secure research environments that prevent data exfiltration) while enabling independent audit studies.

Public bias audit registries: Building on New York City Local Law 144's bias audit requirement, jurisdictions should establish public registries where organizations deploying algorithmic hiring publish audit results. Current NYC implementation allows organizations to publish audits on their own websites, limiting discoverability and comparability. A centralized registry would enable applicants, researchers, and advocates to identify patterns across organizations and vendors.

Standardized evaluation protocols: Regulatory agencies should develop standardized methodologies for evaluating algorithmic hiring systems, analogous to how medical devices undergo FDA review or how vehicles undergo safety testing. Standards would specify appropriate validation methods, fairness metrics, sample size requirements, and documentation expectations. Vendors could submit systems for certification, and employers could require certification as a procurement criterion.

Protected researcher communications: Researchers studying algorithmic hiring often depend on vendor cooperation for data access, creating conflicts of interest that may suppress negative findings. Policy should protect researcher independence through mechanisms like third-party data trusts, preregistration of research designs, or statutory protections analogous to whistleblower safeguards.

Cross-Sector Learning and Knowledge Infrastructure

Algorithmic accountability challenges transcend individual organizations or sectors. Lessons from fairness interventions in one context may inform approaches elsewhere. Effective accountability requires building knowledge infrastructure that enables learning across contexts.

Industry working groups and pre-competitive collaboration: Competing employers may nonetheless share common interests in improving algorithmic hiring quality and fairness. Industry associations can convene working groups to share best practices, develop common evaluation standards, and coordinate vendor engagement. Pre-competitive collaboration of this type exists in other domains (e.g., financial services firms collaborate on anti-money laundering approaches despite competing for customers).

Academic-practitioner partnerships: Universities can provide neutral venues for employer collaboration on algorithmic accountability challenges. HR Analytics research centers at major universities already partner with employers on workforce topics; similar models could focus on algorithmic hiring fairness, validation methods, and alternative approaches. These partnerships could generate published research that informs broader practice.

Government-led knowledge repositories: Regulatory agencies like the U.S. Equal Employment Opportunity Commission could develop clearinghouses of algorithmic hiring research, guidance documents, case studies, and enforcement actions. Making this knowledge accessible to employers, vendors, applicants, and researchers would support evidence-based practice.

International policy coordination: Algorithmic hiring operates across borders, with global vendors serving multinational employers. Policy fragmentation across jurisdictions creates compliance complexity and may enable regulatory arbitrage. International coordination through organizations like the OECD or ILO could harmonize standards, share enforcement approaches, and prevent a race to the bottom.

Rights-Based Frameworks and Worker Voice

Beyond technical governance mechanisms, algorithmic accountability in hiring implicates fundamental questions about worker rights, human dignity, and labor market power. Several emerging frameworks connect algorithmic systems to rights-based analysis.

Algorithmic impact assessments: Drawing from privacy impact assessments and environmental impact reviews, some jurisdictions now require organizations deploying high-risk algorithmic systems to conduct and publish impact assessments (EU AI Act, 2024). These assessments document system purposes, technical approaches, stakeholder consultation, risk mitigation measures, and ongoing monitoring. For hiring algorithms, impact assessments should address employment discrimination risks, labor market concentration effects, and pathways for worker voice.

Applicant rights frameworks: Policy can establish baseline rights for job applicants subjected to algorithmic screening: rights to notification that algorithmic systems are used, rights to meaningful information about assessment criteria, rights to human review of algorithmic decisions, and rights to contest potentially erroneous outcomes. The EU's GDPR already provides some protections (Article 22 limiting fully automated decisions with legal or significant effects), though implementation in employment contexts remains uncertain.

Worker data cooperatives: Individual applicants lack bargaining power to demand algorithmic accountability from employers or vendors. Collective action mechanisms—data cooperatives, applicant advocacy organizations, or labor unions—could aggregate voice and leverage. These organizations might negotiate for transparency, contest systematic disparities, or provide education about applicant rights under algorithmic hiring regimes.

Public employment alternatives: If private sector algorithmic hiring produces systematic exclusion or discrimination, public employment programs could serve as labor market safety valves. Government-funded job creation, job guarantees, or hiring subsidies for excluded populations could mitigate worst-case outcomes while creating pressure for private sector improvement.

Conclusion

This research establishes three core findings about algorithmic monoculture in hiring. First, disaggregated analysis reveals adverse impact masked by aggregate metrics. While vendor-reported aggregate fairness appears acceptable, position-level analysis demonstrates that 10.62% of positions adversely impact Black applicants and 5.32% adversely impact Asian applicants—affecting over 150,000 applications in our dataset. This finding has immediate compliance implications under existing employment discrimination law, which evaluates selection procedures at the position level, not in aggregate.

Second, shared dependence on a vendor's algorithms produces systematically higher rates of homogeneous outcomes than independent decision-making. Of applicants submitting ten applications, 4% face rejection from all positions, and systemic rejection rates decay more slowly than statistical baselines predict. Comparison with the largest prior study of hiring discrimination (Kline et al., 2022) reveals that algorithmic monoculture generates qualitatively different labor market dynamics than traditional processes. Our simulation analysis demonstrates that even broad job searches may not ensure applicants receive favorable evaluations, requiring as many as 25 applications to achieve rejection rates below 0.1%—more than double the baseline of independence.

Third, current approaches to algorithmic accountability inadequately address systemic risks of vendor consolidation. Vendor-conducted bias audits, employer-level fairness reviews, and regulatory frameworks focused on individual organizations all miss the cross-employer, labor-market-level consequences of algorithmic monoculture. When the same algorithms mediate hiring at multiple employers, discrimination becomes structural rather than firm-specific, and individual rejections compound into systemic exclusion.

Actionable implications follow for multiple stakeholders:

For regulators and enforcement agencies: Measure adverse impact at appropriate levels of granularity (per position, not aggregated across vendors or employers). Strengthen market surveillance of algorithmic hiring vendors, potentially requiring transparency reports or bias audit registries. Monitor algorithmic monoculture as a structural labor market concern, not merely a firm-specific compliance issue. Consider mandatory researcher access to vendor deployment data analogous to Digital Services Act provisions for online platforms.

For employers: Disaggregate fairness analysis by position when deploying algorithmic hiring. Implement vendor diversity strategies rather than standardizing on single providers. Maintain meaningful human involvement in hiring decisions rather than delegating entire processes to algorithms. Conduct rigorous validation studies linking assessment performance to employment outcomes rather than accepting vendor claims about general efficacy.

For vendors: Transparently report position-level fairness metrics rather than only aggregate statistics. Provide clients with sufficient documentation to conduct independent validation. Enable algorithmic pluralism by supporting interoperability and data portability rather than locking clients into proprietary systems.

For researchers: Advocate for mandated data access to enable independent audit studies of deployed algorithmic hiring. Develop rigorous methodologies for evaluating labor-market-level consequences of algorithmic monoculture, extending beyond firm-specific analyses. Connect algorithmic fairness research to broader questions of labor market inequality, talent allocation, and worker wellbeing.

The employment relationship remains fundamental to modern life—shaping identity, providing livelihood, and structuring opportunity. As algorithmic systems increasingly mediate access to employment, we must ensure these technologies serve genuine social needs rather than entrenching inequality, enabling discrimination, or creating systemic exclusion. This research demonstrates that current practices fall short of this standard and that substantial reforms are both necessary and feasible.

Research Infographic

Dismantling Algorithmic Monocultures Slide Deck

References

1. Ajunwa, I. (2021). An auditing imperative for automated hiring. Harvard Journal of Law and Technology, 34(1), 621–679.

2. Anderson, E. (2017). Private government: How employers rule our lives (and why we don't talk about it). Princeton University Press.

3. Autor, D., & Scarborough, D. (2008). Does job testing harm minority workers? Evidence from retail establishments. The Quarterly Journal of Economics, 123(1), 219–277.

4. BBC. (2024, February 14). AI hiring tools may be filtering out the best job applicants.

5. Bertrand, M., & Mullainathan, S. (2004). Are Emily and Greg more employable than Lakisha and Jamal? A field experiment on labor market discrimination. American Economic Review, 94(4), 991–1013.

6. Black, E., Koepke, J. L., Kim, P., Barocas, S., & Hsu, M. (2024). Less discriminatory algorithms. Georgetown Law Journal, 113(1), 1–58.

7. Bommasani, R., Creel, K., Kumar, A., Jurafsky, D., & Liang, P. (2022). Picking on the same person: Does algorithmic monoculture homogenize outcomes? Advances in Neural Information Processing Systems, 35, 3933–3946.

8. Bursell, M., & Roumbanis, L. (2024). After the algorithms: A study of meta-algorithmic judgments and diversity in the hiring process at a large multisite company. Big Data & Society, 11(1), 205395172312217.

9. Cowgill, B. (2020). Bias and productivity in humans and algorithms: Theory and evidence from résumé screening. Columbia Business School Working Paper.

10. Equal Employment Opportunity Commission. (1978). Uniform guidelines on employee selection procedures. Federal Register, 43(166), 38290–38315.

11. Farber, H. S., Herbst, C. M., Silverman, D., & von Wachter, T. (2019). Whom do employers want? The role of recent employment and unemployment status and age. Journal of Labor Economics, 37(2), 323–349.

12. Fuller, J., Raman, M., Sage-Gavin, E., & Hines, K. (2021). Hidden workers: Untapped talent. Harvard Business School Project on Managing the Future of Work and Accenture.

13. Hoffman, M., Kahn, L. B., & Li, D. (2018). Discretion in hiring. The Quarterly Journal of Economics, 133(2), 765–800.

14. Hsieh, C. T., Hurst, E., Jones, C. I., & Klenow, P. J. (2019). The allocation of talent and US economic growth. Econometrica, 87(5), 1439–1474.

15. Hu, L. (2023). What is "race" in algorithmic discrimination on the basis of race? Journal of Moral Philosophy, 21(1-2), 1–26.

16. Jacobson, L. S., LaLonde, R. J., & Sullivan, D. G. (1993). Earnings losses of displaced workers. American Economic Review, 83(4), 685–709.

17. Jain, S., Suriyakumar, V., Creel, K., & Wilson, A. (2024). Algorithmic pluralism: A structural approach to equal opportunity. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (pp. 1821–1835).

18. Johnson, G. M. (2020). Algorithmic bias: On the implicit biases of social technology. Synthese, 198(10), 9941–9961.

19. Kassir, S., Baker, L., Dolphin, J., & Polli, F. (2023). AI for hiring in context: A perspective on overcoming the unique challenges of employment research to mitigate disparate impact. AI and Ethics, 3(3), 845–868.

20. Kleinberg, J., & Raghavan, M. (2021). Algorithmic monoculture and social welfare. Proceedings of the National Academy of Sciences, 118(22), e2018340118.

21. Kline, P., Rose, E. K., & Walters, C. R. (2022). Systemic discrimination among large US employers. The Quarterly Journal of Economics, 137(4), 1963–2036.

22. Kroft, K., Lange, F., & Notowidigdo, M. J. (2013). Duration dependence and labor market conditions: Evidence from a field experiment. The Quarterly Journal of Economics, 128(3), 1123–1167.

23. Kuhn, P., Shen, K., & Zhang, S. (2020). Gender-targeted job ads in the recruitment process: Facts from a Chinese job board. Journal of Development Economics, 147, 102531.

24. Nawrat, A. (2023, May 15). Inside HireVue's acquisition of Modern Hire. Unleash.

25. Nonnecke, B., & Carlton, C. (2022). EU and US legislation seek to open up digital platform data. Science, 375(6581), 610–612.

26. Raghavan, M., Barocas, S., Kleinberg, J., & Levy, K. (2020). Mitigating bias in algorithmic hiring: Evaluating claims and practices. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (pp. 469–481).

27. Rhea, A. K., Markey, K., D'Arinzo, L., Schellmann, H., Sloane, M., Squires, P., Khan, F. A., & Stoyanovich, J. (2022). An external stability audit framework to test the validity of personality prediction in AI hiring. Data Mining and Knowledge Discovery, 36(6), 2153–2193.

28. Rothstein, D. S. (2016). An analysis of long-term unemployment. Monthly Labor Review, July 2016.

29. Sen, A. (1997). Inequality, unemployment and contemporary Europe. International Labour Review, 136(2), 155–171.

30. Sloane, M., Moss, E., & Chowdhury, R. (2022). A Silicon Valley love triangle: Hiring algorithms, pseudo-science, and the quest for auditability. Patterns, 3(2), 100425.

31. Sloane, M., Solano-Kamaiko, I. R., Yuan, J., Dasgupta, A., & Stoyanovich, J. (2023). Introducing contextual transparency for automated decision systems. Nature Machine Intelligence, 5(3), 187–195.

32. Sullivan, D., & Von Wachter, T. (2009). Job displacement and mortality: An analysis using administrative data. The Quarterly Journal of Economics, 124(3), 1265–1306.

---

Jonathan H. Westover, PhD is Chief Research Officer (Nexus Institute for Work and AI); Associate Dean and Director of HR Academic Programs (WGU); Professor, Organizational Leadership (UVU); OD/HR/Leadership Consultant (Human Capital Innovations). Read Jonathan Westover's executive profile here.

---

Suggested Citation: Westover, J. H. (2026). Algorithmic Monocultures in Hiring: When One Vendor's Bias Becomes Everyone's Problem. Human Capital Leadership Review, 37(4). doi.org/10.70175/hclreview.2020.37.4.7