Bias by Algorithm: How AI Hiring Tools Discriminate — and Whether New Laws Can Stop Them

The pitch was seductive: replace fallible, prejudiced human recruiters with objective, data-driven AI. No more résumés discarded because a name sounds Black. No more interviewers unconsciously favoring candidates who remind them of themselves. Just clean, efficient pattern-matching against the traits of your best performers. Hundreds of major employers bought in — and millions of job seekers paid the price.

The Promise vs. The Reality

The business case for AI screening was straightforward. With hundreds of applicants per role, human reviewers inevitably rely on shortcuts. Algorithmic tools, vendors promised, would standardize evaluation and remove the idiosyncratic snap judgments that fuel bias. The reality has been considerably more complicated.

Researchers at Stanford and elsewhere have repeatedly demonstrated that AI hiring systems can amplify bias rather than neutralize it. Audit studies using identical résumés with stereotypically white versus Black-sounding names found that algorithmic screeners, trained on historical hiring data, reproduced the same disparate outcomes as their human predecessors — just faster and at greater scale. HireVue, one of the leading vendors of AI-powered video interviews, faced intense scrutiny over its facial analysis features, which critics argued flagged candidates based on expressions and vocal patterns that correlate with race and disability status. The company eventually discontinued facial analysis in 2021, but the broader industry practice of inferring candidate quality from audiovisual signals remains widespread.

The through-line in each documented failure is the same: a system trained on historical outcomes will learn historical prejudices.

Mobley v. Workday Unpacked

No single legal development has rattled the AI hiring industry more than Mobley v. Workday. Derek Mobley, a Black man over 40 who identifies as having anxiety and depression, applied for more than 100 positions at companies using Workday’s AI screening platform — and was rejected for all of them. His class-action lawsuit alleges that Workday’s system systematically discriminates on the basis of race, age, and disability, functioning as an employment agency under federal law and therefore subject to Title VII, the ADEA, and the ADA.

The scale embedded in the complaint is staggering. Workday’s platform is used by approximately 10,000 employers. The putative class of harmed applicants could exceed one billion rejected candidates. When a federal judge certified key portions of the case in May 2025, it sent a clear signal: courts are willing to hold AI intermediaries — not just end-user employers — liable for discriminatory outcomes baked into their products. For an industry accustomed to treating algorithmic decisions as a client’s responsibility, that certification was a watershed moment.

Where Bias Enters the Machine

Understanding why these systems fail requires following bias to its source — and there are several.

Training data problems are the most fundamental. If a model learns to predict “successful hire” from a dataset of past employees, and that workforce was assembled through historically biased practices, the model will encode those biases as predictive signals. It isn’t learning what makes a good employee; it’s learning what your previous managers preferred.

Proxy discrimination is subtler and harder to litigate. Anti-discrimination law prohibits using protected characteristics like race or disability status directly. But AI systems don’t need to see race to discriminate by it. Zip code correlates tightly with race due to decades of residential segregation. Vocabulary choices and sentence structure correlate with educational access, which correlates with socioeconomic background, which correlates with race. Gap years correlate with disability. A model that penalizes any of these features is engaging in proxy discrimination — plausibly deniable, difficult to audit, and potentially illegal under disparate impact doctrine.

Feedback loops compound the problem over time. When a biased system rejects qualified candidates from underrepresented groups, those candidates never enter the workforce data that future models train on. The bias self-reinforces across model generations.

The Regulatory Patchwork

Legislatures have begun responding, but the emerging regulatory landscape is fragmented and uneven.

• New York City Local Law 144 (effective since July 2023) requires employers using automated employment decision tools to conduct annual bias audits by independent auditors and publish summary results. It was the first law of its kind in the U.S. — but enforcement has been tepid, audit methodologies vary wildly between firms, and the law covers only tools used for New York City roles.

• California’s 2025 AI employment rules, enacted under SB 1047’s successor framework, impose transparency and impact assessment requirements on covered employers, but leave significant discretion to companies in defining what constitutes a sufficient assessment. Critics note that self-certification is not the same as accountability.

• Colorado’s AI Act, set to take effect in June 2026, is among the most ambitious state-level frameworks. It requires developers and deployers of high-risk AI systems — explicitly including employment tools — to disclose known risks of algorithmic discrimination, conduct impact assessments, and provide consumers with explanations of consequential decisions. It also establishes a private right of action, which regulators in other states have stopped short of.

What these laws collectively don’t do: mandate technical standards for auditing, require pre-deployment testing, or create a federal floor that applies uniformly across all employers.

What Meaningful Accountability Looks Like

Legal compliance and genuine fairness are not the same thing, and the gap between them is where discrimination currently lives.

Real accountability requires explainability by design — systems that can articulate, in human-auditable terms, why a candidate was rejected. It requires independent auditing standards developed by bodies with no financial stake in the tools being assessed. And it requires disparate impact testing conducted not just on proxy variables, but on actual hiring outcomes disaggregated by race, age, gender, and disability status — published, not buried in compliance filings.

The Mobley litigation may ultimately do more to reshape industry practice than any single statute. If Workday faces liability for outcomes baked into its product, every AI hiring vendor now has a financial incentive to audit aggressively and remediate proactively. That is, at minimum, a better incentive structure than voluntary best practices.

The algorithm was never neutral. The question now is whether the law can be.