What Is Algorithmic Bias in Hiring?
Algorithmic bias in hiring occurs when an AI system used to screen, score, or rank candidates produces systematically different outcomes for different demographic groups, in ways unrelated to actual job-relevant capability. It typically originates not from the algorithm inventing prejudice on its own, but from learning patterns present in the historical data it was trained on. Hunkenschroer and Luetge (2022), in a systematic review of 51 articles on the ethics of AI-enabled recruiting, mapped algorithmic bias as one specific risk within a broader and more balanced picture that also includes genuine opportunities and several questions the research has not yet settled.
Where Algorithmic Bias Actually Comes From
AI recruiting tools are deployed across several stages of hiring, including writing job advertisements, screening resumes, and analyzing video interviews through facial and voice analysis. At each of these stages, an algorithm learns patterns from historical data about who was previously hired, promoted, or rated highly.
If that historical data reflects past human bias, whether in who got interviewed, who got hired, or who received strong performance ratings, an algorithm trained on it can learn and reproduce those same patterns, often without anyone involved intending that outcome. The widely reported case of an internal resume screening tool that one major technology company scrapped in 2018, after discovering it had learned to penalize resumes containing the word "women's," is a well-documented illustration of this mechanism. The tool was trained on a decade of past hiring patterns in a male-dominated industry, and it learned to replicate that imbalance rather than correct for it.
A Newer Risk: Algorithmic Monoculture
Bias does not only come from a single tool's training data. It can also come from many employers using the same vendor, so that one biased system shapes outcomes across the whole market. Researchers call this algorithmic monoculture.
Bommasani, Bana, Creel, Jurafsky, and Liang (2026) ran the largest study of its kind, analyzing 4 million applications from 3.4 million real applicants across 156 employers, all screened by a single vendor's algorithms. They found clear racial disparities at the level of individual positions. Nearly 26 percent of applications from Black candidates, and about 15 percent from Asian candidates, went to positions where the tool adversely impacted their group.
Two findings matter most for HR. First, aggregate fairness checks can look clean while adverse impact still exists position by position, so audits need to happen at that finer level. Second, when many employers rely on the same system, some candidates are rejected everywhere at once, a pattern the researchers describe as systemic rejection.
A More Complete Picture: Opportunities, Risks, and Ambiguities
Hunkenschroer and Luetge (2022) organized their systematic review around three categories, and the resulting picture is more nuanced than a simple story about AI being good or bad for fairness in hiring.
Genuine opportunities identified in the literature include the potential reduction of certain forms of human bias, greater consistency in how candidates are evaluated, faster feedback for applicants, efficiency gains for organizations, and enhanced capacity for recruiters to focus on higher-value work. AI, used well, does not only introduce risk. It can also reduce the inconsistency and unstructured bias that characterizes much human decision-making in hiring.
Genuine risks identified include the introduction of algorithmic bias, loss of candidate privacy and a resulting power imbalance between candidates and employers, a lack of transparency and explainability in how decisions are made, obfuscation of accountability when a decision is attributed to "the algorithm," and the potential loss of meaningful human oversight in consequential decisions.
Genuine ambiguities, meaning questions the research has not yet resolved, include the actual effect of AI recruiting on workforce diversity outcomes, questions about informed consent and the use of candidates' personal data, the impact of AI tools on the validity and accuracy of hiring decisions, and how fair candidates actually perceive AI-driven processes to be. Hunkenschroer and Luetge (2022) treat these as open research questions rather than settled facts in either direction.
Why the Ambiguities Matter as Much as the Risks
It is tempting to treat algorithmic bias as a fully understood, well-quantified problem. The honest picture is more complicated. Hunkenschroer and Luetge (2022) specifically identified the effect of AI recruiting on workforce diversity as unresolved in the literature, meaning the same tool can plausibly reduce bias in one implementation and introduce it in another, depending entirely on the data, design, and oversight involved.
This has a direct practical implication. Organizations cannot assume that deploying AI in hiring automatically makes the process more biased, nor can they assume it automatically makes the process fairer. Both outcomes are possible, and the deciding factor is how carefully the tool was designed, trained, tested, and monitored, not whether AI was involved at all.
How to Reduce Algorithmic Bias Risk
Audit training data for historical bias before deployment. If a model is trained on past hiring or performance data, that data should be examined specifically for patterns that reflect prior discrimination, not assumed to be neutral simply because it is historical.
Test for adverse impact on the tool's actual outputs. Request or conduct adverse impact analyses across demographic groups on the tool's real outputs, using the Four-Fifths Rule or equivalent statistical standards, rather than relying on a vendor's assurance that bias was accounted for during design. As Bommasani et al. (2026) show, check adverse impact position by position, since an aggregate pass can hide disparities within specific roles.
Preserve meaningful human oversight of consequential decisions. Hunkenschroer and Luetge (2022) specifically flag the loss of human oversight and the obfuscation of accountability as genuine risks. Maintaining a clear, documented point of human review and accountability addresses both concerns directly.
Demand transparency and explainability from any AI vendor. A tool whose decision logic cannot be explained or audited should be treated with particular caution, since a lack of transparency is itself one of the specific risks the research identifies, independent of whether measurable bias is ultimately detected.
Monitor outcomes continuously. Algorithmic bias can emerge or shift over time as the underlying population, job requirements, or data distribution changes. A one-time validation at deployment is not sufficient for ongoing assurance.
How Deeper Signals Approaches This
At Deeper Signals, addressing algorithmic bias starts with keeping the underlying measurement grounded in validated, transparent psychometric instruments rather than opaque machine learning models trained directly on unstructured behavioral data. The Core Drivers Diagnostic and Core Values Diagnostic are tested for adverse impact across gender, age, and ethnicity, with all ratios documented and available for review, directly addressing the transparency and accountability risks Hunkenschroer and Luetge (2022) identify.
Sola, the platform's AI assessment assistant, is built with guardrails that prevent it from making employment decisions independently, and its outputs are traceable back to specific, validated assessment constructs rather than functioning as an unexplainable black box. This design directly targets the accountability and transparency risks the research identifies.
Frequently Asked Questions
Does using AI in hiring always introduce bias?
No. Hunkenschroer and Luetge (2022) found that AI can plausibly reduce certain forms of human bias, such as inconsistency in how candidates are evaluated, while also risking the introduction of new, different forms of bias. The outcome depends on the data, design, and oversight involved, not on whether AI is used at all.
How does algorithmic bias actually get introduced into a hiring tool?
Most commonly, through training data that reflects historical human bias. If an algorithm learns from past decisions that favored or disadvantaged certain groups, it can reproduce those same patterns in new decisions, even without anyone intending that outcome.
Is algorithmic bias illegal?
Selection procedures, including AI-driven ones, that produce adverse impact on protected groups are subject to the same legal frameworks as any other selection procedure, including requirements to test for and address disparate impact. Several jurisdictions have introduced additional specific regulation for automated employment decision tools.
Can algorithmic bias be completely eliminated?
Complete elimination is a difficult standard to guarantee for any selection method, human or algorithmic. The realistic goal is ongoing measurement, transparency, and active mitigation, rather than a one-time fix that assumes the problem is permanently solved.
What is the difference between algorithmic bias and human bias in hiring?
Human bias in hiring tends to be inconsistent and varies by individual evaluator. Algorithmic bias, once present in a system, tends to be applied consistently and at much greater scale, which means a biased algorithm can affect far more candidates than a single biased human evaluator, making rigorous testing especially important.
Last reviewed by Dr. Reece Akhtar — June 2026
References
Bommasani, R., Bana, S. H., Creel, K. A., Jurafsky, D., & Liang, P. (2026). Algorithmic monocultures in hiring. In Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency (FAccT '26). https://doi.org/10.1145/3805689.3812400
Hunkenschroer, A. L., & Luetge, C. (2022). Ethics of AI-enabled recruiting and selection: A review and research agenda. Journal of Business Ethics, 178(4), 977–1007. https://doi.org/10.1007/s10551-022-05049-6
