Background

The study was conducted by researchers at the Stanford Institute for Human-Centered Artificial Intelligence (HAI), in collaboration with colleagues at Chapman University and Northeastern University. The authors — Rishi Bommasani, Sarah H. Bana, Kathleen A. Creel, Dan Jurafsky, and Percy Liang — describe it as the first large-scale examination of hiring algorithms operating in real-world conditions. The paper is scheduled to be presented at the ACM Conference on Fairness, Accountability, and Transparency in Montreal.

Researchers tracked approximately 3.4 million applicants who submitted around 4 million applications to roughly 1,700 job postings across 150 employers in 11 industry sectors, covering the period from December 2018 through December 2022. Every application in the dataset was screened by the same third-party AI hiring platform. The data came from talent-assessment vendor pymetrics, which was acquired by Harver in 2022.

In simple terms: the researchers were able to look inside the “black box” of automated hiring at a scale that earlier studies could not, because every application passed through one identifiable system.

What the Study Examined

The study tested two distinct concerns. The first was whether AI screening produces racial disparities in who advances to the next stage of hiring. The second — and the focus of the concentration concern — was what happens to job seekers when one vendor’s model effectively governs the first round of screening across many separate employers.

To measure bias, the researchers applied the U.S. Equal Employment Opportunity Commission’s “four-fifths rule,” a long-standing federal benchmark used under Title VII of the Civil Rights Act of 1964. The rule flags a potential adverse-impact problem when one group is selected at less than 80% of the rate of the most-selected group.

Key Findings

Racial disparities in screening

The researchers reported substantial evidence of racial disparities in AI-based candidate screening. They found that 26% of Black applicants and 15% of Asian applicants applied to at least one position where the AI system disadvantaged their racial group under the four-fifths standard. Measured by applications rather than applicants, just under 26% of applications from Black candidates went to positions flagged for adverse impact.

The study estimated that if the tool had recommended Black and Asian candidates at the same rate as the most-favored group — typically white applicants — roughly 40,000 additional applications from those candidates would have advanced to the next stage of hiring.

The authors stressed that how bias is measured changes the result. Pooling every position together — treating the vendor as a single giant hiring process — made the disparity disappear, because favorable patterns in some job categories canceled out unfavorable patterns in others. Only when each position was analyzed separately, the way employment-discrimination law is actually applied, did the adverse impact become visible.

The “algorithmic monoculture” effect

The study’s second finding speaks directly to the concentration of control. The researchers describe an “algorithmic monoculture” — the condition in which many independent decision-makers rely on the same or similar algorithms. When that happens, a candidate the model declines at one employer tends to be declined everywhere the model is used.

The data bore this out. Applicants who submitted multiple applications to positions screened by the same vendor were more likely to be rejected from every position they applied to than would be expected if each employer were deciding independently. The researchers found that 10% of applicants who submitted four applications screened by the vendor were rejected from all four.

In simple terms: because one company’s software sits between job seekers and many employers, the same individuals — and members of the same racial groups — can be filtered out repeatedly, even when they are applying to entirely different companies.

Independent Decisions Looked Different

To test whether this pattern was specific to shared algorithmic screening, the researchers compared their results against the largest prior study of hiring outcomes, which sent roughly 83,000 applications to 108 Fortune 500 firms over the same period and did not center on AI use.

In that comparison set, the rate at which applicants were rejected from every firm they applied to was no higher than statistical models would predict if each company decided independently. The clustered, repeated-rejection pattern appeared in the shared-vendor data but not in the independent-decision data — evidence the authors say points to market concentration as the driving factor.

Why It Matters for Job Seekers

The timing sharpens the stakes. The study notes that companies are seeing nearly three times as many applications for entry-level positions as in 2022, as a slower entry-level market collides with AI tools that let job seekers apply at scale. In that environment, automated first-round screening determines which applicants a human ever sees.

The authors argue that AI screening tools combine three properties that should not coexist in high-stakes decisions: they are pervasively adopted, highly consequential, and largely opaque to the public. An applicant rejected by a shared model generally has no way to know that the same software is screening them elsewhere, or why it declined them.

Analysis: A Compliance and Concentration Question

The study reframes algorithmic bias as partly a market-structure problem rather than a flaw in any single deployment. When screening for an entire industry concentrates in one vendor, individual employers may not detect that qualified applicants are being filtered out, because each company sees only its own funnel — not the correlated rejections happening across the market.

That distinction also carries legal weight. The researchers emphasize that U.S. discrimination law is applied position by position, while vendors often report fairness metrics in aggregate. A blended, company-wide dashboard can show no disparity even when specific job postings would fail the four-fifths test if examined on their own. Regulatory frameworks moving in this direction include New York City’s Local Law 144, which requires bias audits for automated employment decision tools, and the European Union’s AI Act, which classifies hiring systems as high-risk.

The authors’ central recommendation is structural: the value of, and need for, independent research into algorithmic hiring. Without outside scrutiny, they argue, it will remain difficult to build evidence-based policy governing how these tools shape individual job prospects and the overall composition of the workforce. Harver, which acquired the vendor whose data was studied, did not immediately respond to requests for comment reported by outlets covering the research.

What Recourse Do Applicants Have?

The study documents the problem but does not resolve a question many job seekers now face: what can an applicant actually do after being screened out by an automated tool? Existing rights vary sharply depending on where the applicant lives and whether the employer is private or governmental.

Notice, opt-out, and disclosure rights

At the federal level, there is no single statute that grants private-sector applicants a general right to obtain their algorithmic score. The primary protection remains Title VII of the Civil Rights Act, enforced by the EEOC, which prohibits employment practices that produce unlawful adverse impact — but it operates after the fact, typically through a complaint or charge, rather than by entitling applicants to see the tool’s output.

Local and state laws go further in some jurisdictions. New York City’s Local Law 144 requires employers using an automated employment decision tool to notify candidates at least 10 business days before use, to identify the job qualifications the tool assesses, and to inform candidates of their right to request an alternative selection process or a reasonable accommodation. Under the law, candidates may also request information about the type of data the tool collects, its source, and the employer’s data-retention policy. Employers must additionally post a summary of an independent bias audit conducted within the prior year.

In simple terms: in New York City, an applicant has a right to be told a tool is being used and to ask for a human alternative — but even there, the law does not clearly guarantee a copy of an individual numeric “AI score.”

Enforcement has its own limits. A December 2025 audit by the New York State Comptroller concluded that the Department of Consumer and Worker Protection’s system for enforcing Local Law 144 was ineffective, and a separate field study of 391 NYC employers found that only a small fraction had posted the required audit reports or transparency notices — indicating that rights on paper are not always available in practice.

Should applicants be able to receive their “AI score”?

Whether candidates should be entitled to a copy of their score is an open policy question rather than a settled right. Worker advocates argue that disclosure is necessary to detect and challenge bias, while vendors and some employers contend that scoring methodologies are proprietary trade secrets. That tension — transparency for applicants versus protection of vendor models — sits at the center of current debates and is reflected in newer frameworks such as the European Union’s AI Act, which classifies hiring systems as high-risk, and in pending state-level proposals in the United States.

Is this used for government jobs — and does FOIA apply?

AI is being introduced into public-sector hiring. The U.S. Office of Personnel Management has rolled out an AI tool called USA Class to help generate and classify federal position descriptions, and OPM has signaled broader use of skills-based and AI-assisted screening across federal hiring. OPM officials have stated publicly that the agency is not using AI to make high-level hiring decisions, positioning current tools as support for human reviewers.

Government use is where public-records law becomes relevant. The federal Freedom of Information Act applies to federal agencies — not to private employers — and state equivalents apply to state and local agencies. In principle, an applicant or journalist can file a FOIA or state public-records request seeking records about an algorithmic system a government agency uses, including contracts, audits, and documentation.

In practice, transparency advocates note significant obstacles. Agencies may respond that they do not hold responsive records because a private vendor retains them, or that the requested material is exempt as a proprietary trade secret. Civil-liberties groups that file algorithmic-system records requests advise framing requests broadly enough to capture government-held documentation while anticipating vendor confidentiality claims.

In simple terms: FOIA can be a tool when a government agency is the employer, but it generally reaches the agency’s own records — not a private company’s applicant scores — and vendors often assert trade-secret protections over the underlying scoring.

State Legislation at a Glance

In the absence of a comprehensive federal law, states have moved to regulate AI in hiring directly, producing what legal analysts describe as a patchwork of requirements. As of early 2026, lawmakers in 45 states had introduced more than 1,500 AI-related bills, with hiring and employment among the most active categories. The table below summarizes selected measures affecting how applicants are screened, what they must be told, and what recourse they may have. Status reflects the legislative landscape as of June 2026 and is subject to change.

State	Measure	Status	Key provisions for applicants
New York City	Local Law 144 of 2021	In effect (Jul 2023)	Bias audit required; notice to candidates; right to request an alternative process.
Illinois	HB 3773 (amends Human Rights Act)	In effect (Jan 2026)	Notice when AI is used in employment decisions; bars discriminatory use; draft rules add disclosure and 4-year recordkeeping.
California	FEHA automated-decision rules	In effect (Oct 2025)	Bars discriminatory automated decision systems; anti-bias testing is treated as evidence; 4-year records.
Texas	HB 149 (Responsible AI Governance Act)	In effect (Jan 2026)	Bars intentional discrimination by AI systems; disparate impact alone does not establish intent.
Colorado	SB 24-205 (Colorado AI Act)	Enacted; effective Jun 30, 2026	Duty of reasonable care against “algorithmic discrimination”; narrow small-employer exemption.
Connecticut	SB 5 (AI Responsibility & Transparency Act)	Passed legislature (May 2026)	Would require plain-language disclosure when AI materially influences employment decisions.
Connecticut	SB 435 (ADS Protections for Employees)	Pending	Would add disclosure and notice duties, third-party bias audits, and mandatory human review.
California	SB 947 (“No Robo Bosses” successor)	Pending	Would restrict ADS in employment decisions, require human review, and create a private right of action.
California	SB 951 (Cal-WARN amendment)	Pending	Would expand layoff-notice rules where AI or automation drives displacement.
Texas	HB 1709	Pending	Would require employers to disclose AI use in employment decisions and explain the system’s basis.

Status key: green = in effect; yellow = enacted or passed, not yet effective; orange = pending. NYC Local Law 144 is a municipal ordinance; California’s FEHA provisions are regulations. List is selective, not exhaustive.

In simple terms: where an applicant lives increasingly determines what they are owed. A candidate in New York City, Illinois, or California already has notice or anti-bias protections, while pending bills in California, Connecticut, and Texas would add disclosure, human-review, and in some cases the right to sue.

Conclusion

The Stanford-led study does not claim that AI hiring tools are uniformly broken. It documents something narrower and, in some respects, more difficult to address: when one vendor’s model becomes the gatekeeper for many employers at once, the ordinary assumption that a rejected applicant can simply try elsewhere begins to break down. Both the racial disparities and the repeated, correlated rejections trace back to the same root — a single algorithm making determinations across a market that is supposed to consist of independent decision-makers.

For job seekers, employers, and regulators, the finding suggests that the question is no longer only whether an individual hiring tool is fair, but how much of the hiring market any single tool should be allowed to decide.

Key Takeaways

A Stanford-led study of roughly 4 million applications found that when many employers use the same AI screening vendor, the same candidates can be rejected across multiple companies.
The study found 26% of Black applicants and 15% of Asian applicants applied to positions where the AI system showed adverse impact under the EEOC four-fifths rule.
10% of applicants who submitted four applications screened by the same vendor were rejected from all four — a clustering not seen when companies decided independently.
Aggregated, company-wide fairness reports can hide disparities that appear when each job posting is analyzed separately, as discrimination law requires.
The researchers call the pattern an “algorithmic monoculture” and argue market concentration in hiring software is a systemic risk for job seekers.
Applicant recourse varies by location: there is no general federal right to obtain an “AI score,” but NYC’s Local Law 144 grants notice, opt-out, and limited disclosure rights.
AI is entering public-sector hiring (e.g., OPM’s USA Class tool). FOIA and state public-records laws apply to government employers, but vendor trade-secret claims can limit what is disclosed.
States have created a patchwork: NYC, Illinois, California, and Texas measures are in effect, Colorado’s and Connecticut’s take effect in 2026, and bills in California, Connecticut, and Texas remain pending.

Related Coverage

Editor’s note: Related Coverage links use SEO-optimized anchor text and placeholder URLs. Map each to the corresponding live NexfinityNews article (or publish the topic) before posting; remove any row without a matching article to avoid broken internal links.

Sources

Stanford Institute for Human-Centered AI (HAI), “AI Hiring Tools Can Yield Racial Bias and Systemic Rejection,” May 26, 2026.
Bommasani, Bana, Creel, Jurafsky, Liang, “Algorithmic Monocultures in Hiring,” ACM Conference on Fairness, Accountability, and Transparency (FAccT).
Fortune, “Largest study of AI hiring algorithms to date finds ‘clear racial disparities,’” May 26, 2026.
The Register, “AI hiring algorithms reject Black, Asian job seekers at higher rates,” May 27, 2026.
Marketplace, “AI hiring tools can still have racial biases, study finds,” May 28, 2026.
NYC Department of Consumer and Worker Protection, “Automated Employment Decision Tools (AEDT),” Local Law 144 of 2021.
Office of the New York State Comptroller, “Enforcement of Local Law 144 — Automated Employment Decision Tools,” Dec. 2, 2025.
U.S. Office of Personnel Management, “OPM Expands Access to AI-Powered USA Class Tool,” May 5, 2026.
U.S. Equal Employment Opportunity Commission, Title VII of the Civil Rights Act of 1964 (adverse-impact / four-fifths rule).
Perkins Coie, “Navigating the Growing Landscape of State AI Employment Bills and Laws,” Mar. 4, 2026.
MultiState, “State AI Legislation Tracker 2026,” accessed June 2026.
Reed Smith, “State AI hiring tool regulations filling federal void,” Apr. 22, 2026.
Employment Law Letter, “AI Hiring Tools Under Scrutiny,” May 18, 2026 (Connecticut SB 5 / SB 435)