Skip to main content

AI Hiring Bias: The Evidence — Cited Sources, 2024–2025

|May 16, 2026|Updated Nov 3, 2026|13 min read

Quick Answer: The 2024 University of Washington / AIES study tested three open-source text-embedding models that rank résumés against job descriptions and found White-associated names significantly preferred in 85.1% of 27 tests (Black-associated names in 8.6%) when only the name on otherwise identical résumés changed, across more than 3 million comparisons (Wilson & Caliskan, arXiv:2407.20371). That LLM bias sits on top of an offline baseline measured by Kline, Rose & Walters across 83,000 applications to 108 Fortune 500 employers, a 2.1 percentage-point gap in employer contact between distinctively Black and distinctively white names (NBER WP 29053). The EEOC has already settled one AI-hiring age case (EEOC v. iTutorGroup, $365,000, announced 11 September 2023), and in Mobley v. Workday (N.D. Cal.) the court granted preliminary collective certification on the age (ADEA) claim only in May 2025. As of 7 October 2026 (the date of the latest court order we reviewed), the court has preliminarily certified a collective on the age (ADEA) claim, about 15,000 people have filed consents to join it, the plaintiffs have asked the court to certify a class on race, sex, age and disability claims (hearing set for 9 March 2027), Workday must file its motion to decertify the collective by 10 November 2026, and the merits have not been decided. Each figure on this page links to the source it comes from: papers, a working paper, court records, agency pages, and company, survey and news publications.

A statistic repeated from article to article can lose its source along the way, until a reader can no longer tell which study or filing it came from. This page keeps the link. Every number below links to the source it comes from, and the URL is in the prose, not buried in footnotes. Those sources are papers and a working paper, court records, agency pages, and company, survey and news publications, and each figure is attributed to the one it comes from. The point is not to be comprehensive. It is to let a reader check each figure against the source it comes from.

Claims on both sides of the bias debate, that AI removes bias and that AI is irredeemably biased, are often made without a source. This page lists the studies and regulator filings below, each with its source, and it does not settle the debate.

The Anchor Stat: White Names Preferred in 85.1% of 27 Tests Across 3 Million Comparisons

The strongest causal evidence that AI résumé ranking can be biased comes from the 2024 University of Washington and AIES study by Kyra Wilson and Aylin Caliskan. The paper tested three open-source text-embedding models (E5-mistral-7b-instruct, GritLM-7B and SFR-Embedding-Mistral), which rank résumés against job descriptions and are not chatbots, using 554 résumés, 571 job descriptions and 120 first names. That gave more than three million résumé–job comparisons. The work was NIST-funded and at the AIES 2024 conference. Findings from the paper:

  • White-associated names were significantly preferred in 85.1% of 27 model-by-occupation tests, and Black-associated names in 8.6%. These are shares of tests, not of résumés or comparisons.
  • Gender: on 30 August 2026 the authors revised the paper to say that a bug in their code inverted the gender-only results, so the figures first published for male names (52%) and female names (11%) refer to female and male names respectively. The race and intersectional results are unaffected.
  • White-male names were significantly preferred over Black-male names in all 27 tests.

Primary sources: arXiv preprint 2407.20371, UW News release, Brookings analysis.

The methodology is causal: the résumé text is held constant and only the name is changed, so systematic shifts in the models' rankings come from the name. The authors also report that changing the name frequency-matching strategy "can alter whether Black names or White names are favored in a majority of cases", so the result depends on how names are chosen. The study tested three open-source text-embedding models and does not show that any commercial screening product behaves this way; the authors note that "there are no external gender or racial bias audits of AI-mediated resume screening tools, the majority of which are typically closed-source, propriety, and not accessible for external review".

The Offline Baseline: A 2.1-Point Contact Gap Across 108 Fortune 500 Employers

AI-screening bias does not arrive into a neutral hiring world. It compounds on top of decades of measured human bias. The cleanest contemporary baseline is the Kline, Rose & Walters NBER paper, working paper 29053, published in the Quarterly Journal of Economics. The team sent 83,000+ fictitious applications to 108 of the largest US employers (Fortune 500-scale), real geographically dispersed jobs. Names were randomised to be distinctively white or Black, male or female. Other résumé features were also randomised to control for confounds. Headline results:

  • Distinctively Black names reduced the probability of employer contact by 2.1 percentage points relative to distinctively white names, an effect equal to 9% of the Black mean contact rate.
  • About 20% of firms account for nearly half of the total racial gap. The discrimination is firm-specific, not diffuse. Some Fortune 500 employers can be identified with high statistical confidence as systematic discriminators.

Primary source: NBER WP 29053 abstract, full PDF at nber.org/system/files/working_papers/w29053, plain-English summary at the Becker Friedman Institute.

The older Kang, DeCelles, Tilcsik & Jun study, published in Administrative Science Quarterly (2016), is a field experiment in which the authors responded to 1,600 job postings with résumés that were or were not “whitened”. Black applicants received callbacks on 10.0% of applications with unwhitened résumés and on 25.5% when both the name and the experiences were whitened: a gap of 15.5 percentage points, “a ratio of roughly 2.5 to 1”. For Asian applicants the figures were 11.5% and 21.0% (9.5 points, roughly 1.8 to 1). Half of the postings (800) contained explicit pro-diversity language, and the authors report that “employers that adopt pro-diversity statements are in fact just as likely to engage in discrimination against unwhitened résumés as employers that do not display such statements.” Source: ASQ paper.

For the international comparison: the 2024 Monash / ANU correspondence audit in Australia sent 12,000+ applications to 4,000+ leadership-role job ads. English-named candidates received 26.8% positive responses; non-English-named candidates received 11.3%, a ~57% lower callback rate for identical résumés (Monash Business School). The pattern is global and the magnitudes are large.

Age: iTutorGroup and Mobley v. Workday

Age discrimination has produced the most consequential US legal record on AI hiring to date, because the statutory line is unambiguous (the ADEA protects workers 40 and older) and because the algorithmic patterns are easy to evidence.

EEOC v. iTutorGroup (suit filed 2022, settlement announced 11 September 2023). According to the EEOC's lawsuit, iTutorGroup programmed its tutor application software to automatically reject female applicants aged 55 or older and male applicants aged 60 or older, and rejected more than 200 qualified applicants based in the United States because of their age.

The complaint also alleges that one applicant over 55 was rejected when she applied with her real birth date, then applied again with a later birth date and otherwise identical information and was offered an interview. Settlement: $365,000 to be distributed to applicants who were automatically rejected, plus training, a new anti-discrimination policy and injunctions, monitored by the EEOC for at least five years. Primary sources: EEOC settlement announcement, EEOC suit announcement, and the complaint on the Civil Rights Litigation Clearinghouse.

Mobley v. Workday is a pending AI-hiring action. Filed February 2023, the case has had two early consequential orders, summarised here; later orders are covered in the dedicated explainer linked below. On 12 July 2024, Judge Rita Lin (Northern District of California) found that the complaint plausibly alleges Workday acted as an "agent" of its employer customers, which can make a software vendor to which employers delegate hiring functions liable as an employer under Title VII, the ADEA and the ADA. The court let the disparate impact claims go forward and dismissed the claims that rest on an employment agency theory and the intentional discrimination claims.

On 16 May 2025 the same court granted preliminary collective certification on the age (ADEA) claim only, for people aged 40 and over who, from 24 September 2020, applied for jobs through Workday's application platform and were denied employment recommendations.

As of 7 October 2026 (the date of the latest court order we reviewed), the court has preliminarily certified a collective on the age (ADEA) claim, about 15,000 people have filed consents to join it, the plaintiffs have asked the court to certify a class on race, sex, age and disability claims (hearing set for 9 March 2027), Workday must file its motion to decertify the collective by 10 November 2026, and the merits have not been decided.

The order records Workday's own statement that "1.1 billion applications were rejected using Workday" in that period. It is Workday's figure for all rejected applications, not a count of the collective, and the court said that estimate ignores the qualifiers in the collective's definition. Primary sources: Civil Rights Litigation Clearinghouse case 44074 (full docket and orders), Seyfarth Shaw analysis of the agent ruling.

The full Mobley story, the agent theory, the preliminary collective certification, what discovery will surface, and what jobseekers should document if they fit the class, is covered in the dedicated explainer, Mobley v. Workday, Explained.

Worker Sentiment: AARP's 2024 Survey of 3,580 Workers Age 50+

The legal cases above tell us automated age screening is real. AARP's 2024 Job Change Survey tells us workers know it. The survey, fielded 17 October – 5 November 2024 by NORC AmeriSpeak and Prodege panels with a sample of n = 3,580 US adults aged 50+, weighted to Census benchmarks, was authored by Rebecca Perron and published 15 January 2025:

  • 74% of workers aged 50+ believe their age will be considered a barrier to hiring. 42% call it a major barrier.
  • 34% are concerned AI could specifically jeopardise their job security.
  • 24% plan to change jobs in 2025, up from 14% the prior year. Most cite money, not preference.

Primary source: AARP Public Policy Institute. Read alongside the iTutorGroup settlement and the Mobley litigation, the AARP fear is understandable. Automated screening that rejects older applicants is what the EEOC alleged, and settled, in iTutorGroup.

Socioeconomic Class: The First-Generation Penalty

Race and age dominate the AI-bias conversation. Class gets less attention. One study of first-generation socioeconomic bias in hiring is by Peter Belmi, Margaret Neale, Melissa Thomas-Hunt, and Karina Raz, published in Organization Science (the Stanford GSB summary is dated 20 July 2023). It has three parts:

  1. An audit of 1,783 entry-level job postings, comparing applications that disclosed first-generation-college status with applications that did not.
  2. A survey of 285 hiring managers on their beliefs about the influence of class background.
  3. An experiment with 1,250 people testing whether the way first-generation status is framed changes whether the candidate is considered.

Findings:

  • Stanford GSB reports: “The first-gen applications were significantly less successful, receiving 26% fewer callbacks for an interview than the group that didn't disclose the candidate's background.”
  • 62% of the surveyed hiring managers agreed that “even when individuals have left their original social environment, their behavior is still strongly determined by their social origin” (Stanford GSB wording).
  • In the experiment, among participants primed with a focus on shortcomings, “only 26% said they would consider the first-gen candidate”; among those primed with a focus on strengths, 47% “said they would explore a job offer with the candidate”. These are shares of participants in an experiment, not callback rates.

Primary sources: Stanford GSB Insights (institutional summary) paper at Organization Science 2023.1682.

What this supports: in the audit, applications that disclosed first-generation status received fewer interview callbacks than those that did not, and in the experiment the way the status was framed changed how many participants said they would consider the candidate. The study does not measure any AI screening tool, and we found no source showing that automated hiring software would reproduce the penalty. It does not tell you whether to disclose.

What Employers Themselves Believe

ResumeBuilder ran a survey of 948 business leaders (launched 9 October 2024; page last updated 22 October 2024) that produced one of the most cited self-report figures on this topic. Of companies using AI in hiring, the page says: “9% of companies using the technology report it always produces biased recommendations, while 24% say it often does. Another 34% say bias occurs sometimes, 30% believe it happens rarely, and only 4% say AI never produces bias.” These are business leaders’ opinions, not a measurement of bias. The same page says 51% of companies are leveraging AI in hiring, with 68% expected by the end of 2025, and “Today, 82% of companies use AI to review resumes” (it does not state the base). Primary source: ResumeBuilder.

The gap is the story: leaders say their tools produce bias to some degree and use them anyway, and the survey does not say why. Workday Recruiting customers processed 173 million applications in H1 2024, up 31% year on year, while job openings grew only 7%, applications now scale at roughly four times the rate of openings (Workday Global Workforce Report). That is Workday's count of applications passing through its system; it does not say how many of them an AI screener scored.

Gender, Disability, and the Vendor Track Record

The full bias picture extends beyond race and age. Three cases fill in the rest.

Amazon's experimental recruiting tool (built from 2014). Reuters reported on 10 October 2018, citing five people familiar with the effort, that Amazon's machine-learning team had been building programs since 2014 to review job applicants' résumés. Reuters wrote that “Amazon's computer models were trained to vet applicants by observing patterns in resumes submitted to the company over a 10-year period. Most came from men”, and that the system “penalized resumes that included the word 'women's,' as in 'women's chess club captain.' And it downgraded graduates of two all-women's colleges, according to people familiar with the matter.” According to the article, the team was disbanded “by the start of last year” (2017) and Amazon's recruiters “looked at the recommendations generated by the tool when searching for new hires, but never relied solely on those rankings”. Amazon said the tool “was never used by Amazon recruiters to evaluate candidates”, and Reuters wrote that the company “did not dispute that recruiters looked at the recommendations generated by the recruiting engine”. The account rests on anonymous sources. Source: Reuters, 10 October 2018.

HireVue (2021). In a press release of 12 January 2021, HireVue said that early in 2020 it proactively removed the visual analysis component from all of its new assessments, because its internal research showed that advances in natural language processing had significantly increased the predictive power of language, so visual analysis no longer significantly added value. The same release published an algorithmic audit of one early-career assessment by Cathy O'Neil's firm ORCAA, which concluded that the assessments "work as advertised with regard to fairness and bias issues". Vendor announcement: HireVue press release.

ACLU charges against Aon (late 2023). The ACLU states that in late 2023 it filed EEOC class-wide charges against Aon and an employer that uses Aon's assessments, alleging discrimination based on race and disability on behalf of a biracial (Black/white) autistic job applicant and other similarly situated applicants. These are the ACLU's allegations. On 30 May 2024 the ACLU also filed a complaint with the FTC alleging that Aon deceptively markets its hiring products as "bias-free". Source: ACLU press release, 30 May 2024.

On the regulator side, on 12 May 2022 the EEOC and the DOJ each released a technical assistance document about disability discrimination when employers use AI and other software tools to make employment decisions. The EEOC's release says these tools "may result in unlawful discrimination against people with disabilities in violation of the Americans with Disabilities Act (ADA)". The EEOC's strategic enforcement plan for fiscal years 2024–2028 names AI in hiring as an explicit priority. Sources: EEOC press release, 12 May 2022, DOJ AI & ADA resource.

Three Citation Problems to Watch For

Three patterns to watch when you read statistics on this topic:

1. The "Cornell 20–60%" myth. A widely shared LinkedIn claim asserts that Cornell candidates with manual résumés lose in 20–60% of selection cases. We searched the Cornell ILR School, Cornell Career Services, NBER, arXiv, and Google Scholar and found no such study. A citation without a paper title and DOI cannot be checked. The detailed take-down lives in our sibling article, The Cornell 20–60% ATS Myth, Debunked.

2. The "75% ATS auto-rejection" figure. We found no primary source for it. Fortune wrote in 2014 that it “probably came from a study by a job search services firm called Preptel”, and Preptel’s own site said in August 2013 that it had shut down. The sources that ask directly give different answers: in an Enhancv study, 23 of 25 US recruiters interviewed said their systems do not auto-reject for formatting, content or design, while ResumeBuilder’s survey of business leaders says 21% of companies automatically reject candidates at all stages without human review. None of them measures the overall share. Detail: The 75% ATS Rejection Claim.

3. Vendor surveys as the only anchor. Resume.io, Resume Now, Resume Genius, and ResumeBuilder all sell résumé-related products or services, so their surveys are not independent research. This page cites one of them, ResumeBuilder, once, as a survey of business leaders' self-reports. The discipline is to pair vendor surveys with academic sources (NBER, arXiv) and agency or court records (EEOC, court orders) wherever they exist.

The AI résumé detector claims are also worth flagging here. Originality.ai says its current models (5.0 and 5.0 Max, October 2026) have “99%+ accuracy” and reports its own benchmark results, with sample sizes, in its launch post; we have not reproduced them, and the two tests below were run before that model existed. Scribbr’s test of 12 detectors on 30 short texts (page published 26 February 2024, revised 18 November 2024) scored GPTZero at 52% and Originality.ai at 76%, with partial credit for uncertain answers, so these are small-sample scores, not the share of texts classified correctly; Scribbr also sells an AI detector, which topped its own table at 84%. On non-native English writers, Liang et al. found that seven GPT detectors “classified more than half of TOEFL essays (61.22%) written by non-native English students as AI-generated” (Stanford HAI, arXiv:2304.02819). That was a test on 91 essays, not on résumés. More on the claim: Originality.ai’s 99% accuracy claim: what independent tests found.

What the Evidence Means If You Are the Jobseeker

Five practical conclusions follow from the evidence, in descending order of confidence:

  1. If you are 40 or older and applying through Workday-screened employers, document every rejection. The preliminarily certified Mobley collective covers people aged 40 and over who applied for jobs through Workday's application platform since 24 September 2020 and were denied employment recommendations. It was an opt-in collective, and its opt-in period has closed (the court accepted consent forms filed on or before 14 April 2026). The merits have not been decided as of 7 October 2026. The orders are on the Civil Rights Litigation Clearinghouse docket. Save timestamps, employer names, and the rejection emails.
  2. If you write in English as a second language and worry about AI detectors, keep dated drafts. The one study we found (Liang et al., reported by Stanford HAI) tested seven GPT detectors on 91 TOEFL essays by non-native English students and found that they “classified more than half of TOEFL essays (61.22%) written by non-native English students as AI-generated”. It tested essays, not résumés, and we found no source showing that employers use a detector as a hiring gate.
  3. If you are first-generation college, the Belmi data is mixed news. In the audit, applications that disclosed first-generation status received 26% fewer interview callbacks. In the experiment, participants primed with a focus on strengths were more willing than those primed with a focus on shortcomings to consider the candidate (47% against 26%). The study does not say how to word a disclosure on a real résumé, and it does not measure AI screening.
  4. If you are Black and applying to large US employers, the firm-level discrimination is concentrated. The NBER 29053 abstract reports that firms in the top quintile of racial discrimination were responsible for nearly half of the contacts lost by Black applicants. A careers-page diversity statement is not a guide to how an employer treats applicants: in the Kang et al. field experiment, employers that adopted pro-diversity statements were “just as likely to engage in discrimination against unwhitened résumés” as employers that did not display such statements.
  5. Verify the role fits before optimising the résumé. A polished AI-assisted résumé for the wrong role still fails. Run a structured assessment, Career Match for role shortlist, our methodology page for how the matching works, and the AI in Hiring guide for the broader funnel context. The full résumé-side playbook for working with AI without getting flagged: How to Write a Résumé with AI Without Getting Rejected. The companion citation hub: AI Résumé Statistics 2026.

Headline Evidence, Source URLs in One Table

#FindingSource (n)Source URL
1White names preferred in 85.1% of 27 tests by three text-embedding models (gender results corrected by the authors, August 2026)UW & AIES (3M+ comparisons)arxiv.org/abs/2407.20371
22.1-point employer-contact gap for Black names across Fortune 500Kline/Rose/Walters NBER (n=83,000)nber.org/papers/w29053
3Black applicants: 10.0% callbacks with unwhitened résumés, 25.5% with name and experiences whitened (1,600 postings)Kang et al., ASQ 2016journals.sagepub.com/doi/abs/10.1177/0001839216639577
4EEOC alleged iTutorGroup software auto-rejected 200+ US applicants on age; $365K settlement (announced Sept 2023)EEOC enforcementeeoc.gov/newsroom/itutorgroup-pay-365000-settle-eeoc-discriminatory-hiring-suit
5Mobley v. Workday: preliminary ADEA collective certification only (May 2025); merits not decided as of 7 October 2026NDCA case 3:23-cv-00770 (Dkt. 128)clearinghouse.net/case/44074/
674% of workers 50+ see age as a hiring barrierAARP (n=3,580)aarp.org/pri/topics/.../older-job-seekers-age-discrimination-artificial-intelligence
7First-gen disclosure: 26% fewer interview callbacks; in an experiment, 26% vs 47% of participants would consider the candidate, by framingBelmi et al., Stanford GSB / Org Sciencegsb.stanford.edu/insights/do-first-gen-college-grads-face-bias-job-market
89% of business leaders say AI hiring tools always give biased recommendations, 24% oftenResumeBuilder (n=948)resumebuilder.com/7-in-10-companies-will-use-ai-in-the-hiring-process-in-2025-despite-most-saying-its-biased
911.3% callback for non-English names (vs 26.8% English) in AU leadership rolesMonash / ANU (n=12,000+)monash.edu/news/articles/study-confirms-english-sounding-names-get-more-call-backs
10EEOC and DOJ each released a document (12 May 2022): AI hiring tools may result in unlawful disability discrimination under the ADAFederal technical assistance, May 2022eeoc.gov/newsroom/us-eeoc-and-us-department-justice-warn-against-disability-discrimination
11Seven GPT detectors flagged 61.22% of 91 TOEFL essays by non-native English students as AI-generated (essays, not résumés)Liang et al., reported by Stanford HAI (n=91 essays)hai.stanford.edu/news/ai-detectors-biased-against-non-native-english-writers

What This Page Doesn't Claim

For symmetry: the evidence above does not support the claim that every commercial AI screener is equally biased, that all hiring algorithms produce identical disparate impacts, or that human screening would be unbiased in their place. The Kline–Rose–Walters paper supports the second point: it reports that firms in the top quintile of racial discrimination account for nearly half of the contacts lost by Black applicants. The Kang et al. experiment shows that different signals (name only, experiences only, both) produced different callback rates in its sample. And the human-screening results above (Kline–Rose–Walters, Kang et al., the Monash / ANU audit) show that human screening has been measured to be biased too.

The honest summary is narrower. The strongest causal evidence on this page is that three research embedding models ranked identical résumés differently when only the name changed; that human employers contacted distinctively Black names less often in the NBER audit, with a minority of firms accounting for a large share of the gap; that the EEOC alleged, and settled, an age-screening case; and that Mobley v. Workday is being tested in court, where, as of 7 October 2026, the merits have not been decided. None of these sources measures bias in a commercial AI screening product in use today. Whether your specific application was screened out for the right reasons or the wrong ones is not something these sources can tell you for any single case.

A note on dates. The sources on this page carry their own dates. The Mobley preliminary collective certification (May 2025), the AARP survey (January 2025) and UW/AIES (October 2024) are recent; the Kline, Rose & Walters working paper dates from 2021 and the Reuters report on Amazon from 2018. Litigation status is as of 7 October 2026, the date of the latest court order we reviewed. If a link below no longer works, look for an archived copy of the same page rather than a vendor blog.

For teams & organizations
Free assessments for HR teams

281 tests for hiring, onboarding, and team development. No per-employee fees, no certification.

See HR team plans

Ready when you are

Find your ideal career match in 3 minutes.

12 questions. Full result with strengths, blind spots, and careers matched to your type.