Industries / Finance and business services / Staffing and recruiting
Fairness
Making sure that hiring tests, recruiters and software do not shut people out because of race, sex, age or other protected traits, and checking whether they do.
7
Milestones achieved, from 1971 to 2026.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1971
Tests must fit the job
Title VII of the US Civil Rights Act of 1964, in force from 2 July 1965, bars employers from refusing to hire, and employment agencies from refusing to refer, anyone because of race, color, religion, sex or national origin. In Griggs v. Duke Power Co., decided on 8 March 1971, the Supreme Court held that a high school diploma rule and two aptitude tests that screened out Black workers at higher rates, and had not been shown to predict job performance, were unlawful even without intent, because the Act forbids practices "fair in form, but discriminatory in operation."
Credit
US Supreme Court, in a case brought by Black employees of Duke Power's Dan River plant
1978
The four-fifths rule
The Uniform Guidelines on Employee Selection Procedures of 25 August 1978 say that a group selected at less than four-fifths (80%) of the rate of the most-selected group "will generally be regarded" by US federal enforcement agencies as evidence of adverse impact. Smaller gaps can still count if they are significant, and larger ones may not if they rest on small numbers. Nearly all the bias audits of hiring software posted under New York City's 2023 law reported results above the same 0.8 line.
Credit
US Equal Employment Opportunity Commission and federal enforcement agencies (29 CFR Part 1607)
2000
Blind auditions, then doubts
Most top US orchestras began auditioning players behind a screen in the 1970s and 1980s. Claudia Goldin and Cecilia Rouse (American Economic Review, 2000) estimated that the screen raised by 50% the chance that a woman advanced out of some preliminary rounds, and explained 30% to 55% of the rise in women among new hires; women were under 5% of players in the top five orchestras in 1970 and 25% in 1997. In 2019 Jonatan Pallesen and the statistician Andrew Gelman pointed out that many of the paper's key estimates had large standard errors and were not statistically significant, as the authors themselves had noted.
LessonA widely retold result about fair hiring rested on noisy estimates, and the caveat was in the paper's own tables.
Credit
Claudia Goldin and Cecilia Rouse; critique by Jonatan Pallesen and Andrew Gelman
2004
Same résumé, different name
Marianne Bertrand and Sendhil Mullainathan answered help-wanted ads in Boston and Chicago with fictitious résumés given very white-sounding or very Black-sounding names, and found that white names got 50% more callbacks (American Economic Review, 2004). A 2017 meta-analysis of all available US hiring field experiments by Lincoln Quillian and three co-authors found no change in discrimination against African Americans since 1989. In 2021 Patrick Kline, Evan Rose and Christopher Walters sent 83,000 applications to 108 of the largest US employers and found that Black names cut the chance of contact by 2.1 percentage points, with the most discriminating fifth of firms responsible for nearly half the lost contacts.
Credit
Marianne Bertrand and Sendhil Mullainathan; Lincoln Quillian, Devah Pager, Ole Hexel and Arnfinn Midtbøen; Patrick Kline, Evan Rose and Christopher Walters
2018
Amazon drops résumé model
Reuters reported in October 2018 that Amazon had been building software since 2014 to rate job applicants' résumés, trained on résumés sent to the company over 10 years, most of them from men. The software penalized résumés containing the word "women's" and downgraded graduates of two all-women's colleges, and Amazon disbanded the team by the start of 2017; people familiar with the project said recruiters looked at its rankings but never relied on them alone.
LessonA model trained on a company's past hiring learned the imbalance in it.
Credit
Reuters, reporting on Amazon
2023
New York requires bias audits
From 5 July 2023, New York City's Local Law 144 bars employers and agencies from using an automated employment decision tool unless it has had a bias audit within the past year, a summary is posted and candidates are notified. In a study by Lucas Wright, J. Nathan Matias and colleagues (FAccT 2024), 155 student investigators checked 391 employers and found 18 posted audits and 13 notices; because employers decide whether their tools are covered, a missing audit "cannot be said to indicate non-compliance." In December 2025 the New York State Comptroller called the city's enforcement "ineffective": 75% of test calls about such tools to the 311 hotline were misrouted, and the city found one problem in 32 posted audits where the Comptroller found at least 17.
LessonA disclosure law that let employers decide whether it applied to them produced few disclosures.
Credit
New York City Council and Department of Consumer and Worker Protection; Lucas Wright, J. Nathan Matias and colleagues; New York State Comptroller
2026
A deployed screener measured
Rishi Bommasani, Sarah Bana, Kathleen Creel, Dan Jurafsky and Percy Liang studied 4 million applications by 3.4 million people to 1,700 job postings at about 150 employers, all scored by one vendor's AI, and found that 26% of Black applicants and 15% of Asian applicants applied to positions where the tool failed the four-fifths test against their group. The gaps showed only when each position was checked on its own; pooled across all of the vendor's work, the tool showed no adverse impact. Stanford's Institute for Human-Centered AI calls it the first large-scale study of hiring algorithms in the wild.
LessonChecking a hiring tool's results in total can hide discrimination that shows up job by job.
Credit
Rishi Bommasani, Sarah Bana, Kathleen Creel, Dan Jurafsky and Percy Liang (Stanford, Chapman and Northeastern universities)
Watch for
Workday age case proceeds
Derek Mobley claims that Workday's AI tools for recommending job applicants screened out people over 40. On 16 May 2025 Judge Rita Lin of the US District Court for the Northern District of California let the claim go forward as a nationwide collective action, centered on "whether Workday's AI recommendation system has a disparate impact on applicants over forty." On 29 May 2026 the court ruled that Workday did not have to hand over its customers' applicant data, and that its internal bias-testing data was protected by attorney-client privilege.
Achieved whenA court rules on the evidence whether the recommendations had a disparate impact on applicants over 40.
Claimed by
Derek Mobley, collective action certified May 2025
EU hiring AI rules wait
The EU's AI Act (Regulation 2024/1689 of 13 June 2024) lists as high-risk any AI system used "to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates". Its duties for such systems were due to apply from 2 August 2026, but the Digital Omnibus on AI (Regulation 2026/1744 of 8 July 2026) moved the date to 2 December 2027, citing late standards and late national authorities.
Achieved whenThe duties apply and a national authority reports checking hiring systems against them.
Claimed by
European Parliament and Council of the European Union, July 2026
Colorado rewrites its AI law
Colorado's 2024 AI Act, signed on 17 May 2024, would have required companies that build or use high-risk AI systems, including for hiring, to take reasonable care against algorithmic discrimination from 1 February 2026; a 2025 law pushed the date to 30 June 2026. Before it took effect, a law signed on 14 May 2026 repealed it and put in its place rules for automated decision-making technology from 1 January 2027: notice to people when such tools are used, a plain-language explanation within 30 days of a decision against them, and human review on request.
Achieved whenThe 2026 act takes effect and the attorney general adopts its rules.
leg.colorado.govleg.colorado.govleg.colorado.govourtake.bakerbotts.com
Claimed by
Colorado General Assembly, May 2026
Open
01
Name-swap tests of hiring software
A public, repeatable test that runs matched made-up résumés, identical except for names and other signals of race, sex or age, through the commercial screening products employers buy, with their makers' agreement, and through general AI models set up as screeners. Results are reported job by job, using the four-fifths rule, not pooled.
The briefCredit · unclaimed02
A known-answer test for audits
A set of made-up hiring outcome tables, listing applicants, groups, positions and who was selected, with gaps of known size planted in them: gaps that vanish when positions are pooled, gaps that appear only for combined groups such as Black women, and small samples. Auditors and audit software analyze the tables, and then the answer key is published.
The briefCredit · unclaimed
Open 01 · the brief
Name-swap tests of hiring software
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public, repeatable test that runs matched made-up résumés, identical except for names and other signals of race, sex or age, through the commercial screening products employers buy, with their makers' agreement, and through general AI models set up as screeners. Results are reported job by job, using the four-fifths rule, not pooled.
Why it is still open
Published name-swap tests score general AI models, not named hiring products. Kyra Wilson and Aylin Caliskan (2024) found text-embedding models favored white-associated names in 85.1% of cases; Zhenyu Gao, Wenxi Jiang and Yutong Yan (June 2026) found that 14 language models released from 2024 on showed no gap or a gap favoring Black names; Kevin Webster (2025) tested eight AI platforms without naming them. The Stanford-led team says it is the first and only group to study deployed hiring algorithms independently at scale and that data access holds such research back, and in May 2026 a court held Workday's internal bias tests privileged.
arxiv.orgarxiv.orgarxiv.orgalgorithmichiring.github.iocdflaborlaw.com
Built with what exists
Made-up résumés written with Claude or open-weight models and checked by recruiters; name lists from published audit studies; trial or research accounts from vendors; Python and the four-fifths calculation in 29 CFR 1607.4.
First step
Write 20 matched résumé pairs for two job ads, run them through two open-weight models, and report the results job by job.
Done looks like
At least three commercial products and five general models, each run on 50 job ads with 20 matched résumé pairs per ad, with impact ratios per job and per group published with the code, and a repeat run six months later.
Open 02 · the brief
A known-answer test for audits
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A set of made-up hiring outcome tables, listing applicants, groups, positions and who was selected, with gaps of known size planted in them: gaps that vanish when positions are pooled, gaps that appear only for combined groups such as Black women, and small samples. Auditors and audit software analyze the tables, and then the answer key is published.
Why it is still open
New York State's Comptroller found that the city reviewed 32 posted bias audits and found one problem where the Comptroller found at least 17. The Stanford-led study found that pooling one vendor's decisions hid adverse impact that showed up position by position. The nearest work found is a public set of 1,730 synthetic CVs for testing ranking systems (Jorge Saldivar, Anna Gatzioura and Carlos Castillo, 2025), and FairAware (Anna Verheyden and colleagues, August 2026), which planted age, gender and language gaps in 10,000 made-up candidates and asked 70 people, half of them HR staff, to find them with one dashboard. Neither scores professional auditors or audit software, and neither plants gaps that vanish when positions are pooled.
Built with what exists
Python with NumPy or the Synthetic Data Vault to generate the tables; Fairlearn and IBM's AI Fairness 360 as baseline checkers; the four-fifths rule in 29 CFR 1607.4; a GitHub repository.
First step
Build five tables, including one where the pooled impact ratio is above 0.8 while two positions fall below it, and run Fairlearn on them.
Done looks like
30 tables with a published answer key, analyzed by at least two audit firms or audit tools and by the two open-source toolkits, with a table of planted gaps found, gaps missed and false alarms for each.