Algorithmic bias in hiring: how AI affects who gets jobs and what it means

01How AI is used in hiring and recruitment

Artificial intelligence has become deeply embedded in modern recruitment. Companies use AI to scan resumes for keywords, rank candidates against job descriptions, conduct automated video interviews that analyze facial expressions and speech patterns, and predict which applicants are most likely to succeed or stay. By some estimates, over 80 percent of large employers use some form of automated screening in their hiring pipeline.

The appeal is efficiency. A single AI tool can process tens of thousands of applications in hours, a task that would take human recruiters weeks. The promise is objectivity — removing human bias from decisions by letting data decide. The reality is that the data itself carries bias, and the systems trained on it can reproduce and amplify those patterns at scale.

02Where algorithmic bias enters the process

Bias can enter at multiple stages. Training data is the most common source: if a company's past hires were predominantly male, a model trained on that history learns to associate maleness with success. Feature selection matters too — a zip code can serve as a proxy for race, a commute distance can proxy for socioeconomic status, and gaps in employment can correlate with caregiving responsibilities that disproportionately affect women.

Model design choices also contribute. The choice of target variable — what the model predicts — determines what the system optimizes for. If the target is past hiring outcomes, the model reproduces past decisions. If the target is employee tenure, it may favor candidates from privileged backgrounds who can afford to stay in lower-paying roles longer. Each choice encodes a value judgment about what matters.

03The types of bias in AI hiring tools

Algorithmic bias in hiring takes several forms. Representational bias occurs when training data underrepresents certain groups — a model trained mostly on male engineers will underpredict female candidates. Algorithmic bias describes systematic and repeatable harmful tendency in a computerized sociotechnical system to create "unfair" outcomes, such as "privileging" one category over another in ways that may or may not be different from the intended function of the algorithm. Measurement bias arises when proxies for the desired trait systematically misrepresent a group — using college prestige as a quality signal disadvantages qualified candidates from less-prestigious institutions.

Aggregation bias emerges when a single model is applied to diverse populations. A personality assessment calibrated on one demographic may produce lower scores for another, not because those candidates are less capable, but because the assessment's norms do not reflect their communication styles or cultural backgrounds. The result is systematic disadvantage that is invisible to the system's operators.

04Real cases of discriminatory AI hiring

The most prominent case involved Amazon's experimental resume screening tool, which was scrapped in 2018 after it was found to systematically downgrade resumes containing the word 'women's' — as in 'women's chess club captain' — and penalize graduates of all-women's colleges. The model had been trained on ten years of resumes submitted to the company, which were overwhelmingly from men, and it learned to favor male candidates.

Other cases include video interview platforms that were found to rate candidates differently based on skin tone and accent, and personality assessments that screened out candidates on the autism spectrum. These cases are not anomalies; they are predictable consequences of training systems on historical data that reflects existing inequalities and deploying them without adequate testing for disparate impact.

05What regulators are doing about it

Regulatory responses are accelerating. The European Union's AI Act classifies AI systems used in employment as high-risk, subjecting them to mandatory risk assessments, logging, human oversight, and transparency requirements. In the United States, the Equal Employment Opportunity Commission has issued guidance that automated decision tools are subject to existing employment discrimination law, including disparate impact analysis.

New York City's Local Law 144 requires annual bias audits of automated employment decision tools and mandates that candidates be notified when AI is used in evaluation. Illinois, Colorado, and other states have followed with their own requirements. The regulatory landscape is fragmenting across jurisdictions, creating compliance challenges for companies that hire nationally or internationally.

06How companies can audit their AI hiring

Auditing is the primary defense. Companies should test their tools for disparate impact before deployment and regularly thereafter, measuring selection rates across protected groups and applying the four-fifths rule — a legal standard requiring that the selection rate for any protected group be at least 80 percent of the rate for the highest-selected group.

Beyond statistical testing, audits should examine the full pipeline: training data provenance, feature engineering choices, model interpretability, and the human override process. Vendor claims of fairness are not sufficient. Companies retain legal liability for discriminatory outcomes regardless of whether they built the tool or bought it, making independent validation essential.

07What job seekers should know

Job seekers should know that AI may be screening their application before any human sees it. Optimizing for the algorithm — using exact keywords from the job description, formatting resumes in machine-readable text rather than images, and avoiding unusual formatting — can improve the odds of passing automated filters. But this shifts the burden onto candidates to game a system they cannot see.

In jurisdictions with notification requirements, candidates have the right to know when AI is used and, in some cases, to request human review. Understanding these rights — and exercising them — is becoming an essential part of navigating the modern job market, especially for candidates from groups historically disadvantaged by automated screening.