The Bias Problem in AI Hiring: What the Research Actually Shows

LLMs favored white-associated names 85% of the time when ranking identical resumes. That is not a hypothetical. It is the finding of a University of Washington study presented at the AAAI/ACM Conference on AI, Ethics, and Society in October 2024. And it is just one of several independent research efforts converging on the same conclusion: AI hiring tools carry measurable, documented bias.

69% of HR professionals now use AI to support recruiting, up from 51% in 2024 (SHRM 2025 Talent Trends). 42% of companies have implemented AI-powered applicant tracking systems. 41% of recruiters use AI daily for candidate sourcing and screening. The tools are everywhere. The question is what happens when those tools carry the biases of the data they were trained on - at a scale no human recruiter could replicate.

AI bias in hiring: what three studies found

The evidence is not anecdotal. Three major independent research efforts paint a consistent picture.

The University of Washington study (October 2024) tested how LLMs rank resumes with identical qualifications but different names. White-associated names were favored 85% of the time. Female-associated names were favored only 11% of the time. Black male-associated names were never favored over white male-associated names. The study also found that AI bias increases when resumes are shorter - with less content, demographic signals like names carry more weight.

The Brookings Institution research (2024) confirmed that leading AI models systematically disadvantage Black male applicants even when qualifications are identical. The critical finding: biases operate intersectionally. The penalty for being Black and male is not the sum of a race penalty and a gender penalty - it is a distinct, compounding effect that simple bias audits may miss entirely.

A large-scale experiment testing approximately 361,000 fictitious resumes across five LLMs - including GPT-3.5 Turbo, GPT-4o, Gemini 1.5 Flash, Claude 3.5 Sonnet, and Llama 3-70b - found that models systematically favored female candidates while disadvantaging Black male applicants, even with identical qualifications (VoxDev, 2025). Black women faced different outcomes than Black men or white women. The root cause, according to researchers: reliance on historical data reproduces existing racial, gender, and intersectional disparities.

Three studies. Multiple AI models. Hundreds of thousands of resumes. The same pattern. This is not a bug in one system. It is a structural feature of how these tools are built.

The lawsuits that changed the conversation

Research documents the problem. Lawsuits are testing whether the legal system will address it.

Mobley v. Workday is the first AI hiring discrimination case to reach collective action status. Derek Mobley, a Black applicant over 40, applied to more than 100 jobs using Workday's screening system and was rejected nearly every time. Four additional plaintiffs submitted hundreds of applications through Workday with similar results. In May 2025, a federal judge certified the case as a nationwide collective action under the Age Discrimination in Employment Act. The court noted that over one billion applicants may have been rejected using Workday's tools and ordered the company to turn over its client list of employers using its HiredScore screening product.

One billion. That is the scale of automated hiring decisions that are now subject to legal scrutiny.

The Eightfold AI class action (filed January 2026) takes a different legal approach. The suit alleges that Eightfold scraped personal data on more than one billion workers, scored each applicant on a 0-5 scale, and discarded low-ranked candidates before any human review. The legal theory: Eightfold acted as a consumer reporting agency without complying with the Fair Credit Reporting Act - no notice, no accuracy requirements, no dispute rights. The case was brought by former EEOC chair Jenny R. Yang and the nonprofit Towards Justice. Eightfold, used by Microsoft, PayPal, and other major employers, denied the allegations.

Two lawsuits. Two legal theories - disparate impact and FCRA transparency. Two billion people potentially affected. I first examined how AI screening tools filter applications. The bias dimension makes the stakes fundamentally different.

The regulatory fracture

The federal government and state governments are moving in opposite directions.

The EEOC withdrew its AI-related hiring guidance from its website on January 27, 2025. By September 30, 2025, the agency had ceased investigating claims based solely on disparate impact discrimination and closed all pending charges. A Trump executive order in April 2025 abandoned disparate impact as a federal enforcement theory.

The critical caveat: federal anti-discrimination laws - Title VII, the ADEA, the ADA - still apply. The laws did not change. The enforcement posture did. Employees can still bring lawsuits. The Mobley case proves it.

States are filling the vacuum.

Illinois (effective January 1, 2026): Employers cannot use AI in ways that discriminate, regardless of intent. They must provide advance notice to candidates including the AI product name, the decisions it affects, and the data it collects.

Colorado (effective June 2026): The first comprehensive state AI law. Requires impact assessments, worker notification, appeal rights for AI employment decisions, and public disclosure of AI systems used.

New Jersey (effective December 15, 2025): Regulations governing disparate impact of automated employment decision tools.

New York City (Local Law 144): Requires annual bias audits, public disclosure, and 10 business days advance notice before using automated employment decision tools. But a New York State Comptroller audit in December 2025 found enforcement "ineffective" - 75% of 311 calls about the law were misrouted, and the enforcement agency identified only one violation while auditors found 17.

Over 107 new BIPA class action lawsuits were filed in Illinois in 2025 alone, with landmark settlements including Clearview AI ($51.75 million) and Speedway ($12.1 million). The legal landscape is expanding even as federal enforcement contracts.

The result: the AI bias protections available to you depend on where you live.

What job seekers can do

The structural nature of AI hiring bias means individual tactics cannot solve a systemic problem. But they can improve your position.

Know your state's protections. If you are in Illinois, Colorado, New York City, or New Jersey, employers must disclose AI use in hiring decisions and in some cases provide appeal rights. In Illinois, you have the right to know which AI product is screening you and what data it collects. Use these rights.

Understand that shorter resumes amplify bias. The UW research found that bias increases when resumes contain less content. Demographic signals like names carry proportionally more weight when there is less professional content to evaluate. A detailed, keyword-rich resume gives the algorithm more signal to work with beyond your name.

Ask about the screening process. In jurisdictions with disclosure requirements, you can ask whether AI tools are used in the hiring process. Even where disclosure is not legally required, the question itself signals awareness - and may prompt a more transparent response.

Document patterns. If you are applying through the same platform and experiencing consistent rejection despite strong qualifications, the pattern may be relevant to ongoing legal actions. The Mobley v. Workday collective is actively seeking participants who applied through Workday's system.

The ATS systems that filter your applications are being augmented - and sometimes replaced - by AI tools with their own biases. In tech hiring, where application volumes are highest, AI screening is most prevalent. Understanding how these systems work is the first step toward navigating them.

JobIntel's credibility scoring evaluates the listings themselves - not the candidates. The methodology is transparent, the signals are disclosed, and the scoring does not use applicant demographics. When I built the system, building it without the bias was the minimum viable requirement.

---

Try JobIntel free at jobintel.com. See which listings are real before you apply.