Diversity hiring: what actually works, what doesn’t, and what’s now legally risky

TL;DR

• Diversity hiring means designing recruitment so that job-relevant merit, rather than demographic proxy, determines who advances. 
• The problem it addresses is real and measured: field experiments in the US, UK and Australia find large, persistent callback penalties triggered by a name alone. 
• The business case usually quoted for it is far weaker than the industry claims, and the consultancy statistic everyone cites has failed replication twice.
• Meanwhile the interventions most commonly recommended run from the single best-evidenced finding in personnel selection to practices the US enforcement agency is now actively prosecuting.

So this article grades nine of them, on one consistent scale: robust, promising but thinly evidenced, mixed, and unsupported. It also splits the legal position four ways, because advice that is lawful in Sydney or Brussels can be a Title VII exposure in Ohio.

Why this article is different

You are not new to this. 

You have a diversity hiring commitment, a list of twenty recommended recruitment strategies, a legal team that has grown noticeably more nervous since 2025, and no reliable way to tell which items on the list actually move the number and which will get you sued for trying.

Almost every article on this topic presents its list of best practices as equally valid.

This one does two things they don’t: it grades every intervention by the quality of the evidence behind it, and it separates the law by jurisdiction rather than assuming everyone reads from the US playbook. Where our own commercial category has a weak evidence base, we say so.

For context on where we write from: Sapia.ai has delivered more than 10 million candidate interviews and publishes a third-party audit of its scoring models, which we discuss with its limitations below.

This is general information, not legal advice. The legal position in several of these jurisdictions is moving quickly. Take advice before changing policy.

What diversity hiring actually means (and what it doesn’t)

Diversity hiring is the practice of removing artificial barriers and demographic noise from recruitment so that job-relevant signal, not proxy, determines who gets hired. It covers how roles are written and advertised, where you source, how candidates are assessed, how scores are combined, and how you monitor the outcome by stage. 

The terms diversity recruiting, diversity recruitment, diversity staffing, recruiting for diversity, recruitment for diversity, inclusive recruitment, inclusive hiring and diversity and inclusion in recruitment are used interchangeably with it and mean the same thing. Put simply, the diversity hiring meaning comes down to one design choice: what information a selection decision is allowed to rest on.

What hiring for diversity is not is quota or preference hiring. That distinction is this article’s spine, and it is also the legal fault line:

• Diversity hiring acts on the process. It changes what information a decision is based on, and how consistently that decision is made. It is lawful in every jurisdiction covered here, and it has evidence behind it.
• Quota or preference hiring acts on the decision. It uses a protected characteristic – race, sex, disability, age, religion or belief, sexual orientation, and in UK law marital status and civil partnership – as an input into an individual selection outcome. It is narrowly conditioned in the UK, the EU and Australia, and it is currently an enforcement priority in the United States.

Most content on diversity in hiring blurs the two, which is why inclusive hiring practices and preference schemes are so often defended, or attacked, as though they were the same thing. Most of what has become legally dangerous sits in the second category; most of what actually works sits in the first.

The problem is real – here is the evidence for its size

The strongest evidence in this entire field comes from correspondence studies: researchers send matched fictitious applications to real vacancies, varying only the applicant’s name. Every other consideration is held constant, so any gap in employer contact is attributable to the name. It is the closest research design available to observing what two otherwise identical job seekers actually experience.

United States

Patrick Kline, Evan Rose and Christopher Walters, writing in the Quarterly Journal of Economics (2022), sent more than 83,000 fictitious applications to 108 Fortune 500 firms. Distinctively Black names reduced employer contact by 2.1 percentage points, a 9% relative reduction. The more useful finding is that discrimination is concentrated, not diffuse: firms in the top quintile accounted for nearly half of all lost contacts to Black applicants, and the Gini coefficient of employer contact gaps was around 0.4, comparable to US household income inequality. Contact gaps were also considerably smaller at firms with more centralised points of contact. A top-five economics journal, in other words, found that centralised, consistent process is associated with less discrimination.

United Kingdom

Valentina Di Stasio and Anthony Heath at Nuffield College’s Centre for Social Investigation (2019) sent roughly 3,200 fictitious applications to real British vacancies. Minority-ethnic applicants had to send 60% more applications than white British applicants to receive one positive response. On the trend, Heath and Di Stasio’s meta-analysis in the British Journal of Sociology (2019) pooled eleven British field experiments across five decades and concluded that they “can confidently reject the hypothesis that [discrimination] risks have declined,” with the penalty facing Pakistani applicants virtually unchanged over forty years.

Australia

Alison Booth, Andrew Leigh and Elena Varganova, in the Oxford Bulletin of Economics and Statistics (2012), sent over 4,000 fictitious CVs. Callback rates were 35% for Anglo-Saxon names, 26% for Indigenous, 22% for Middle Eastern and 21% for Chinese. To achieve equal callbacks, Chinese applicants needed to send 68% more applications, Middle Eastern applicants 64%, and Indigenous applicants 35%.

One further finding is worth sitting with. Kang, DeCelles, Tilcsik and Jun (Administrative Science Quarterly, 2016) found Black applicants received 25% callbacks with “whitened” résumés versus 10% without – and that employers with explicit pro-diversity statements were no less likely to discriminate, while minority candidates were half as likely to whiten when applying to them. Diversity messaging changed candidate behaviour without changing employer behaviour. –

A note on the business case

The claim that a more diverse workforce makes a company more profitable is the most-quoted argument for workplace diversity, and it has not held up.

McKinsey’s Diversity Matters Even More (2023) reported that companies in the top quartile for ethnic diversity on their executive teams were 39% more likely to financially outperform. 

Jeremiah Green and John Hand, in Econ Journal Watch (2024), attempted a quasi-replication on S&P 500 data and found 37 of 40 test statistics statistically insignificant, with a correlation of r = 0.02. 

Bermiss, Green and Hand then ran 180 forward-looking regressions and found no reliable predictive relationship. 

There is also a design problem: McKinsey measured diversity after the financial performance window it correlates with, so the causal arrow, if there is one, points the wrong way. The underlying company-level data has never been disclosed, so the finding cannot be independently checked.

The peer-reviewed answer is less dramatic and more interesting. Lukas Wallrich and colleagues, in a pre-registered meta-analysis in the Journal of Business and Psychology (2024) covering 615 reports and 2,638 effect sizes, found average relationships between team diversity and team performance that are positive but insubstantial (|r| < .1), moderated by task complexity and creativity. 

Pletzer and colleagues (PLOS ONE, 2015) found board gender diversity effects of r = .01, not statistically significant, and drew the right conclusion: diversity demonstrably does not harm performance.

The same caution applies to the softer version of the argument, that a diverse team fosters innovation because varied perspectives improve decision making. Wallrich and colleagues do find the relationship is moderated by task complexity and creativity, so there is something real there. But it is a small, conditional effect, not the step-change the consultancy literature describes. If enhanced innovation is the plank your programme is standing on internally, it is a weaker one than you have been told.

Which means the case for diversity hiring does not need a profit claim at all. It rests on the fact that measurable discrimination in selection is a validity problem – you are systematically failing to identify capable people – and a legal one. That reframe is what the rest of this article is built on.

How we graded the evidence

What counts as evidence here: 

Pre-registered meta-analyses and randomised or framed field experiments outrank observational panel studies; those outrank single studies; those outrank lab experiments with student participants; and all of them outrank vendor research and consultancy reports. 

Critically, a practice is graded on whether it has been shown to change actual hiring outcomes – not attitudes, intentions or self-reported learning. That distinction alone reclassifies several very popular interventions. Deliberately out of scope: retention, progression, organisational culture and the wider work of building an inclusive work environment once people are hired, plus pay equity. Those are separate literatures with separate obligations.

The five criteria, applied to every intervention below:

1. Evidence grade. Robust (replicated, high-quality design), promising but thinly evidenced (sound and lawful, but supported only by observational data or by studies showing the practice is rarely implemented as claimed), mixed (credible evidence in both directions), or unsupported (popular, without traceable or replicated support).
2. Effect on representation. Does it demonstrably change who gets hired, and by how much? Not who gets interviewed, and not what reviewers report thinking.
3. Effect on predictive validity. Does it hold or improve your ability to predict who performs and stays? A diversity intervention that costs validity gets reversed within two years.
4. Adverse impact profile. The measured subgroup difference (Cohen’s d) where one exists, and whether the practice has been tested for differential prediction rather than just selection rates.
5. Legal exposure in 2026. Lawful and encouraged everywhere, jurisdiction-specific, or high-risk – split by US, UK, EU and Australia where they diverge.

Correcting the numbers everyone is using

Before the interventions, a correction that almost no article in this category has caught up with, and that changes the economics of the entire debate.

Nearly every vendor deck and blog post still quotes Schmidt and Hunter’s 1998 validity coefficients. Those numbers are now known to be overcorrected. Sackett, Zhang, Berry and Lievens showed in the Journal of Applied Psychology (2022), with a practitioner-facing follow-up in Industrial and Organizational Psychology (2023), that decades of meta-analyses applied range-restriction corrections using artefact distributions that overstated restriction. Correcting for it deflates almost every coefficient and reorders the rankings.

The consequence is the point. Because general mental ability’s validity was overstated, the cost of not leaning on it has collapsed. Under the old estimates, removing GMA from a six-predictor composite cut validity by 0.20, from 0.66 to 0.46 – prohibitive. Under the revised estimates it costs 0.05, from 0.61 to 0.56. Sackett and colleagues put it plainly: “equally valid composites with smaller group differences can be created, thus aiding in reducing the validity-diversity tradeoff.”

Read that back against the argument you have probably heard in a leadership meeting. The “you must choose between quality of hire and diversity” trade-off that has justified a decade of inaction is roughly a quarter as expensive as the industry believes. And the predictor that now comes out best on both axes simultaneously is the structured interview.

What actually works – nine interventions, graded

Robust – do these first

Only two interventions earn this grade, and that is itself a finding: the two best-evidenced practices in diversity hiring are also the two least fashionable ones.

1. Structured interviews

Evidence grade: robust – twelve meta-analyses, reviewed systematically. Levashina and colleagues (Personnel Psychology, 2014) called the structured-over-unstructured advantage “one of the most consistent findings in the history of research on the employment interview.” 

What they found: anchored rating scales raise validity from r = .26 to .35 and interrater reliability to .77. The adverse-impact result matters more: pre-1996 studies, largely unstructured, showed d = +0.20 against minority and female candidates; post-1996 structured studies showed d = -0.02, effectively parity. 

In practice: job analysis first, same questions in the same order, behaviourally anchored scales, scored independently before discussion.

Where it falls short: structure is a design property, not a format – the same paper found video interviews in the literature were significantly less structured than face-to-face ones (4.28 versus 6.50 components, d = .96), so buying a video tool does not buy structure. And it only holds if interviewers are prevented from overriding it. 

Legal: lawful and affirmatively encouraged in all four jurisdictions.

2. Scoring criteria mechanically rather than holistically

Evidence grade: robust – meta-analytic, plus large-scale firm data. 

Kuncel and colleagues (Journal of Applied Psychology, 2013) found mechanical combination of assessment data predicted job performance at r = .44 versus .28 for expert holistic judgement, a more than 50% improvement, with the loss occurring “even when the judges are knowledgeable experts.” Hoffman, Kahn and Li (QJE, 2018), across 265,648 hires at 15 firms, found hiring managers overrode job-test recommendations 22% of the time, and a one-standard-deviation increase in override rate was associated with roughly 7% shorter tenure. 

In practice: fix the weights before you see the candidates, log every override, review the patterns quarterly. 

Where it falls short: the QJE paper does not report the demographic composition of hires, so it is not evidence of a diversity effect. And mechanical scoring only helps if the inputs are job-relevant – it will faithfully mechanise a bad criterion. 

Legal: lawful everywhere; documented, consistently applied scoring is also the practical answer to Ames-style claims from any direction.

Sound in principle, but weaker than you have been told

3. Removing artificial barriers

Evidence grade: mixed. Unnecessary degree requirements, over-broad criminal-history screens, narrow experience proxies buried in job descriptions, and application processes that do not work with assistive technology are all worth removing. 

The practice is legally unimpeachable and logically sound. But the only large-scale evidence on whether removal changes hiring says it mostly doesn’t. Sigelman, Fuller and Martin at the Burning Glass Institute and Harvard Business School (2024) examined over 11,000 postings: roles dropping degree requirements nearly quadrupled over a decade, but fewer than 1 in 700 new hires actually resulted, and 45% of firms were practising “skills-based hiring in name only.” Where firms genuinely followed through, non-degree hires showed 10 percentage points higher retention than credentialed peers. 

In practice: announcing it does nothing. Change the screening rule and the requisition template, not the job advert. Write job requirements as what the role genuinely needs on day one. And offer reasonable adjustments proactively at every stage rather than waiting to be asked: in the UK that duty is anticipatory under the Equality Act 2010, and under the ADA it attaches to the application process itself, not only to employment. 

Where it falls short: the study is observational, not causal, and it measures postings against hires, so it establishes an implementation gap rather than an effect size. Barrier removal also expands the pool without telling you how to evaluate it, which is a different problem. 

Legal: lawful everywhere. Be precise rather than reassuring, though: the US Department of Justice has been reinterpreting Title VII disparate impact, and these are exactly the practices it flags for review, from the opposite direction. The defensible version is the affirmative one the DOJ itself recommends – job-focused, consistently applied employment practices.

4. Broad outreach and funnel expansion

Evidence grade: promising, but observational only. Frank Dobbin and Alexandra Kalev’s three-decade panel of over 800 US firms associates targeted college recruitment with roughly 10% increases in women in management positions, and minority-focused recruitment with about 8% more Black male and 9% more Black female managers. 

These are not the largest effects in their dataset: among the inclusion initiatives they tracked, diversity task forces (+9-30%), mentoring programmes (+9-24%) and dedicated diversity managers (+7-18%) all show larger associations. Recruitment sits at the lower end of the positive range. 

In practice: add channels that reach diverse talent – job boards, recruitment agencies, community and university partnerships, outreach programmes – and measure the change in pool composition rather than the effort spent. One channel deserves separate scrutiny, because it is usually the largest and is almost never audited: employee referrals. Referral schemes are the cheapest source of hire most organisations have, and they reproduce the demographic composition of your existing employees, since people refer from their own networks – which is precisely why they rarely surface candidates from diverse backgrounds you do not already employ. If referrals are a large share of your funnel, they are the most likely reason a talent pool stays homogeneous while everything upstream of it changes. 

Where it falls short: this is longitudinal observational panel evidence, not a randomised design, and firms that invest in outreach differ in unobserved ways – every figure here is an association, not an effect. And, again, outreach expands the pool without improving how you evaluate it. 

Legal: outreach is not an employment decision and excludes nobody, which makes it the most defensible lever available. 

But note that Executive Order 14398’s definition of prohibited “racially discriminatory DEI activities” expressly reaches recruitment, and EEOC leadership has publicly questioned the practice of encouraging specific groups to apply. That is an agency-head view and an enforcement risk, not settled law – it is more aggressive than any court has gone. US readers should keep outreach strictly additive (more channels, never a substituted or restricted one) and keep records showing selection itself was characteristic-blind.

Mixed – the evidence cuts both ways

5. Blind or anonymous CV screening

Evidence grade: mixed, with real negative findings. We market a blind screening capability, so here is the evidence straight. Three rigorous studies: one negative, one negative at shortlisting, one genuinely mixed. 

Hiscox and colleagues at BETA, the Australian government’s behavioural economics team (2017), ran a framed field experiment with over 2,100 public servants across 14 agencies and found that removing de-identification made female candidates 2.9 points more likely to be shortlisted and Indigenous female candidates 22.2 points more likely; the authors concluded de-identification “does not appear to assist in promoting diversity.” Behaghel, Crépon and Le Barbanchon’s randomised trial in France (AEJ: Applied, 2015) found participating firms became less likely to interview and hire minority candidates under anonymous résumés. Åslund and Nordström Skans (ILR Review, 2012) found Swedish anonymisation raised interview rates by about 8 points for non-Western immigrants and about 8 points for women – and at the offer stage women gained about 6.5 points, a real and durable effect, while “no such effect is found for immigrants.” It worked for gender; the ethnicity gain did not survive into offers.

In practice: in each case, blinding removed reviewers’ ability to use identity information benignly as well as prejudicially, and it does nothing about the stages after the screen. 

The distinction that matters: these studies test subtracting information from human CV review. They do not test adding a measured, job-relevant signal. Those are different interventions with different logic, and conflating them would be convenient for us and wrong. 

Where it falls short: the BETA study was a hypothetical shortlisting task with volunteer participants and possible demand effects, and the French firms had self-selected as minority-friendly. 

Legal: lawful everywhere.

One more thing worth naming, because it is the most-cited “proof” that blind evaluation works. The claim that blind orchestra auditions increased women’s success by 50%, attributed to Goldin and Rouse (2000), does not appear in that paper. Andrew Gelman’s re-reading of how the study is cited notes that the authors themselves state some key coefficients “do not pass standard tests of statistical significance.”

6. Algorithmic and AI screening

Evidence grade: mixed, and design-dependent – this is not one intervention. Algorithmic screening is the source of both the best and the worst evidence in this field. Li, Raymond and Bergman (“Hiring as Exploration,” Review of Economic Studies) studied roughly 90,000 applications at a Fortune 500 firm. 

A conventional supervised-learning model selected about 2% Black and under 5% Hispanic applicants for interview; an exploration-based model that placed value on informational uncertainty selected 14% Black and 10% Hispanic – and achieved a 25% hire rate among those interviewed, versus 10% for human recruiters. 

This was not demographic targeting: the gain came from valuing candidates the historical data knew least about. Against that, Wilson and Caliskan (AIES, 2024) tested embedding models on over 500 real résumés across nine occupations and found white-associated names favoured in 85.1% of cases, female-associated names preferred in only 11.1%, and Black male candidates disadvantaged in up to 100% of cases, with document length and name frequency driving selection.

Amazon’s abandoned 2018 tool, which learned to penalise the word “women’s,” remains the canonical demonstration that a model trained to reproduce past hiring reproduces past composition. 

“AI in hiring” is not one thing. The variable is what the model is trained to optimise, and whether it measures something job-relevant or infers from history. 

Legal: see below – this is where nearly all the new binding law sits.

7. Diverse slates and diverse interview panels

Evidence grade: mixed to weak. The “two in the pool” finding (Johnson, Hekman and Chan, 2016) is quoted everywhere as established: 79 times greater odds of hiring a woman with two or more women in the candidate pool, 194 times with two or more minority candidates. Treat it carefully. It was published in Harvard Business Review, not a peer-reviewed journal; no independent replication has been located in either direction; and odds ratios of that magnitude in hiring research almost always indicate a near-empty cell rather than a real effect. The best field evidence on slate rules is discouraging: Wade, Rider, Swaminathan and Schwab’s study of roughly 1,300 NFL coaches (1985-2015) in the American Journal of Sociology found the Rooney Rule did not achieve its aim, because white coaches were nearly twice as likely to reach coordinator – the role accounting for about 79% of head-coach promotions – so the bottleneck sat upstream of where the rule applied. 

Where it falls short: the headline finding is unreplicated magazine research, and slate rules act on a stage that may not be where your loss occurs. 

Legal: diverse-slate policies and diverse hiring panels are both named EEOC enforcement targets in the June 2026 National Enforcement Plan. And pretending to comply carries separate exposure: the $85m Wells Fargo settlement approved in 2026 concerned “sham” interviews conducted to satisfy diverse-slate guidelines.

Unsupported – stop spending money here

8. Unconscious bias training

Evidence grade: unsupported for behaviour change. Forscher and colleagues (Journal of Personality and Social Psychology, 2019) meta-analysed 492 studies and 87,418 participants: effects on implicit measures are weak (|d| < .30), the procedures produced “trivial changes in behavior,” and – the decisive result – changes in implicit measures did not mediate changes in behaviour. 

Dobbin and Kalev’s panel data associates mandatory diversity and inclusion training with declines, including roughly 9% fewer Black women in management over five years, while voluntary training was associated with 9-13% gains. In December 2020 the UK Government concluded there is “no evidence that this training changes behaviour in the long term or improves workplace equality” and phased unconscious bias training out of the Civil Service. 

To be fair about the nuance: Bezrukova and colleagues (Psychological Bulletin, 2016) do find real training effects, with an overall Hedges’ g = .38. But the “behavioural learning” measured there is largely in or near the training context – role-plays and self-report – not downstream hiring decisions, and attitudinal effects decay. Training produces learning. It does not produce different hires. The wider lesson runs through this entire list: nothing here will eliminate bias inside individual heads. The interventions that work change what bias is able to act on. More on the distinction in our piece on unconscious bias training.

9. Gendered job-ad language scanning

Evidence grade: unsupported – the outcome was never measured. Gaucher, Friesen and Kay (JPSP, 2011) is the foundation of an entire commercial tool category. Its archival studies found masculine wording differences of d = 0.40 and 0.24, but never measured application behaviour or hiring outcomes. 

Its experimental studies used between 43 and 118 undergraduate participants and measured stated appeal on a Likert scale, with effect sizes of roughly 4-10% of variance. 

Where it falls short: the causal chain runs from wording to more female applicants to more women hired, and only the first link was ever tested – the rest is inferred. No field test of hiring outcomes has been located. A replication and extension (Seong et al., Strategic Entrepreneurship Journal, 2024) exists and is worth reading before drawing conclusions in either direction. Grade it as cheap and harmless rather than effective. Write job adverts in clear, inclusive language because clarity is worth having on its own terms; do not expect it to move your funnel.

Where the law now stands – and why it depends where you are

Start with the sentence that does the most work here, because it is verifiable from two primary sources:

The same mechanism – preferring a candidate from one of the underrepresented groups where two candidates are equally qualified – is legally mandated for large listed EU company boards as of 30 June 2026 under Directive (EU) 2022/2381, and is a named EEOC enforcement priority in the United States as of 4 June 2026.

Three distinctions that most content collapses:

1. Government DEI policy versus private-employer law. Most 2025-26 US activity binds federal agencies and contractors, and Title VII itself has not changed. But enforcement posture and private litigation appetite have both moved sharply, which changes real risk without changing black-letter law.
2. Race-conscious decisioning versus inclusive process. Every jurisdiction still permits – and several now mandate – broad outreach, structured job-related assessment, barrier removal and outcome monitoring. What has become dangerous is using a protected characteristic as an input to an individual decision, or setting numerical outcomes that operate as a proxy for doing so.
3. The jurisdictions are diverging. Advice that is compliant in Sydney or Brussels can be a Title VII exposure in Ohio.

United States

SFFA (2023) was not a Title VII case and did not ban affirmative action in hiring, but its reasoning has been imported into employment enforcement. On 30 June 2026 the EEOC voted to rescind its 1979 Affirmative Action Guidelines (29 CFR Part 1608) and Compliance Manual Section 607. Ames v. Ohio Dept. of Youth Services (June 2025, unanimous) removed the heightened “background circumstances” burden for majority-group plaintiffs, and Muldrow (2024) lowered the harm threshold to “some injury.” In EEOC v. The New York Times (filed May 2026) the agency pleads the employer’s own published representation goals as evidence of discriminatory intent. Two precision points that most commentary gets wrong. First, United Steelworkers v. Weber (1979) and Johnson v. Transportation Agency (1987) have not been overruled: narrowly-drawn voluntary affirmative action plans by private employers remain formally permitted. But an employer relying on Weber or Johnson in 2026 is relying on live precedent that no federal agency will support and that the enforcing agency has publicly repudiated – a different and more accurate statement than “it is now unlawful.” Second, disparate-impact liability is statutory (42 U.S.C. § 2000e-2(k), codified by the Civil Rights Act of 1991) and has not been abolished. Only Congress or the Supreme Court can remove it. Federal enforcement has retreated; private and state exposure has not.

United Kingdom

Positive action under s.158 of the Equality Act 2010 and the narrow s.159 tie-breaker are lawful; positive discrimination is not, and nothing in 2025-26 changed that. Public bodies and many of their suppliers carry an additional obligation that private employers do not: the public sector equality duty at s.149 requires due regard to the need to eliminate discrimination and to advance equality of opportunity, which in recruitment means being able to produce the analysis, not just assert the intention. The Employment Rights Act 2025 brings equality action plans for employers with 250+ staff. Mandatory ethnicity and disability pay gap reporting was confirmed in the March 2026 consultation response but still requires the Equality (Race and Disability) Bill. The operative regulator on automated screening is the ICO: its 2026 report and draft guidance found many employers did not realise they were making solely automated decisions, and insists that human involvement be “active and genuine,” not “a token step.” That is the most practically useful regulator guidance available in any jurisdiction.

European Union

Employment AI remains high-risk under Annex III of the AI Act. Be precise about what changed: the Digital Omnibus (Regulation (EU) 2026/1744, in force 27 July 2026) moved the Annex III compliance deadline to 2 December 2027. It did not reclassify or weaken it. Deployers – that means employers – carry Article 26 duties; Article 27 fundamental-rights impact assessments do not apply to ordinary private employers. Separately, the Pay Transparency Directive’s transposition deadline passed on 7 June 2026 with only a handful of member states complete: pay ranges before interview, no salary-history questions, first reports by 7 June 2027.

Australia

The most permissive of the four. Special measures under s.7D of the Sex Discrimination Act 1984 and its equivalents permit targeted advertising, identified positions and preferential consideration, subject to a four-part test – and with no certification or safe harbour, so employers self-assess. NSW is the outlier and generally needs an exemption. Employers with 500+ staff must now select three targets from WGEA’s Gender Equality Targets Menu, at least one numeric and CEO-signed. Note carefully what those are: organisational targets with a compliance consequence, not selection quotas applied to individual candidates.

One closing observation, because it runs against the obvious reading. Workforce-composition disclosure is contracting on both sides of the Atlantic at once – the EEOC has proposed rescinding EEO-1, and the EU’s Omnibus I Directive cut CSRD scope to employers above 1,000 staff and €450m turnover, sharply reducing the population reporting workforce diversity under ESRS S1. Meanwhile pay disclosure is expanding, through the EU Pay Transparency Directive, UK ethnicity and disability pay gap reporting, and Australia’s employer-level WGEA publication. The comparable data is disappearing at exactly the moment the obligation to explain pay gaps grows.

How to tell whether your assessment is actually fair

The four-fifths rule, and its two caveats

A selection rate below 80% of the highest group’s rate is generally regarded as evidence of adverse impact (29 CFR 1607.4(D)). But the regulation itself says smaller differences can constitute adverse impact where they are statistically and practically significant, and larger ones may not where the numbers are small. Passing four-fifths does not prove the absence of adverse impact.

The elephant: the Uniform Guidelines are under attack 

A DOJ Office of Legal Counsel opinion of 9 June 2026 concluded that the EEOC’s Title VII disparate-impact interpretive guidance, including 29 CFR Part 1607 itself, is unconstitutional, and OFCCP’s August 2026 rules rescind its own adoption of the Guidelines at 41 CFR Part 60-3. Here is the defensible line: an OLC opinion binds executive agencies, not courts; Part 1607 remains on the books unless and until rescinded through rulemaking; and the underlying statutory framework is untouched. The Guidelines remain the right operational discipline and the right evidentiary yardstick – they are simply no longer a safe harbour anyone will defend for you.

Adverse impact is not bias

This is the distinction vendors blur most often. Following Nancy Tippins’ 2023 EEOC testimony on behalf of SIOP, separate measurement bias (systematic error unrelated to the construct), differential prediction (different predictor-criterion regression lines across groups – the one that matters legally), and differential validity. A perfectly valid test can produce adverse impact. Differential prediction is the analysis most often skipped.

What a bias audit is actually worth

Wright and colleagues’ “Null Compliance” study (ACM FAccT, 2024) checked 391 NYC employers under Local Law 144: 18 had posted a bias audit, and 13 had posted a transparency notice. Nearly all published audits reported an impact factor above 0.8 – unsurprising, since the audited party chooses the scope, the comparison groups and the data. The NY State Comptroller’s December 2025 audit found the regulator had received two complaints in two years. Treat weak enforcement as a temporary condition.

The questions to ask a vendor

What was the audit’s scope, and who chose it? Which subgroups, and what sample size? Was differential prediction tested, or only selection rates? Are model versions documented and revalidated? Was the model trained on your current employees or on your actual applicant population? Does the assessment work with screen readers and other assistive technology, and how does a candidate request a reasonable adjustment without that request reaching the people scoring them? And who is liable if it discriminates?

The answer to that last one is: you are

Mobley v. Workday established in July 2024 that a screening vendor can be liable as an agent of the employer, reasoning that “an employer cannot escape liability for discrimination by delegating traditional functions, like hiring, to a third party” – and the employer’s own liability remains fully intact. California’s FEHA automated-decision-system regulations, effective 1 October 2025, contain a comparable agent definition. As of 2026 the Mobley ADEA collective has been conditionally certified and expanded, with opt-in notice authorised in February 2026.

One honest warning

Bias-testing output is evidence. Under California’s regulations it can be evidence in your favour; in litigation it can be evidence against you. Generate it deliberately, and where available, under privilege. That is more useful to know than the pretence that audits are cost-free.

What to measure

Measure the funnel, not the headcount. Most diversity metrics fail for the same reason: a single headline representation number tells you that you have a problem without letting you identify gaps, because it cannot say which stage lost them. Clear diversity metrics are stage-level ones. Analyse data by stage, by group and against an external benchmark.

• Stage-by-stage pass-through rate and adverse impact ratio by group, with a significance test. Aggregate representation hides where the loss happens – and per Kline, Rose and Walters, the loss is concentrated rather than spread evenly. In the UK and Australia the underlying data usually already exists as equal opportunities monitoring; the gap is almost never collection, it is that nobody breaks it down by stage.
• Applicant pool composition against a labour-market availability benchmark. This is the metric that separates a sourcing problem from a selection problem, a distinction most diversity hiring content conflates entirely. If the pool matches availability but your hires don’t, the problem is your process. If the pool doesn’t match, no process fix will help. The disciplined version is OFCCP availability-analysis logic – treat it as analytical method rather than live obligation, since the underlying US contractor requirements were rescinded by OFCCP’s August 2026 rules, but the logic transfers to any jurisdiction.
• Offer-acceptance rate by group. Equal offer rates with unequal acceptance is a candidate-experience or employer-brand problem, not a selection one. Moon, Park and Park’s meta-analysis in Current Psychology found technology-mediated interviews depressed applicants’ behavioural intentions (ρ = -.11 overall), with asynchronous formats much worse (ρ = -.20, versus +.03 for synchronous), and AI-based formats scoring worse on organisational attractiveness than non-AI ones. Modality choices show up in acceptance, not in selection rates – which is also why candidates respond differently to chat and video formats.
• Withdrawal and drop-off by group and stage. This is the plausible place modality effects land, and it deserves an honest statement of the evidence gap: no rigorous study establishing differential withdrawal or completion rates for one-way video interviews by disability, neurodivergence or ESL status has been located. The mechanism is plausible and the ADA exposure is real, but the measured findings above are about attractiveness and intentions, not withdrawal. Measure it in your own funnel rather than borrowing someone’s statistic.
• 90-day and 12-month retention by group, and differential prediction against the performance criterion. The real test of whether widening selection cost you quality. On the available evidence it doesn’t: the Burning Glass/HBS work found non-degree hires at genuine adopters retained 10 points better than credentialed peers.

There is no authoritative regulator-issued metric set for the downstream measures. The anchors are the Uniform Guidelines and SIOP’s Principles together with its January 2023 AI assessment recommendations. Most of what else you will find on this topic is a vendor listicle.

Where Sapia.ai fits (and where it doesn’t)

Sapia.ai‘s AI chat interview is a mobile-first, untimed, text-based AI structured interview with explainable scoring. The reason to consider it follows from the evidence above rather than from any adjective: structured interviews are now the top-ranked single predictor at .42 and carry the lowest subgroup difference of the high-validity predictors (d = 0.23), with post-1996 meta-analytic subgroup differences at d = -0.02. 

To be precise, integrity tests show a lower subgroup difference (d = 0.10) at materially lower validity – so the claim is the best combination of validity and adverse impact, not the lowest subgroup difference outright. The table is printed above; check it. The mechanism is adding a measured, job-relevant signal at the point of application, which is a different intervention from the CV-anonymisation studies above.

The proof

• Stage-by-stage representation held: LNER maintained 30% ethnic-minority representation through each hiring stage, alongside a 57% decrease in time to hire. This is the best proof point on this page precisely because it is a funnel metric rather than a headline number.
• Pool shift: Woodies in Ireland reported three times more ethnic-minority candidates and 1.5 times more women in a three-month period. Their HR manager David Nally put it more bluntly than we would: “It’s done more to help us achieve our D&I goals than all our unconscious bias training.”
• Third-party fairness testing: BLDS, LLC found no evidence of practically significant disparate impact across 23 standardised-mean-difference tests and 49 adverse-impact-ratio tests on models deployed in the US and Canada.
• Completion at scale: Woolworths recorded 82.6% chat interview completion and a 9.2/10 candidate experience score across roughly 27,000 hires in under 10 weeks, running alongside SuccessFactors.

Where Sapia.ai falls short

This article has just told you to interrogate vendor audits, so we have to survive our own test.

1. The BLDS audit is vendor-commissioned, covers models deployed in the US and Canada, and reports SMD and AIR – selection-rate measures. It is not a published differential-prediction study, which is the analysis this article names as the most important one. What we do publish sits on Sapia Labs and our science page.
2. Sapia.ai does not fix pool composition. If your applicant pool doesn’t match labour-market availability, a fairer assessment cannot manufacture diverse candidates who never applied. That is a sourcing and employer-brand problem, and it is often the binding constraint.
3. Sapia.ai is not an ATS or a system of record. It overlays your existing stack via integrations.
4. It is built for volume, frontline, graduate and professional hiring at scale. For a handful of senior or executive hires, neither the economics nor the mechanism fit.
5. It does not address pay-gap reporting, progression or retention interventions – separate problems, separate literatures, and increasingly separate legal obligations in the UK, EU and Australia.

Where to choose something else

Deep technical or coding screening: HackerRank, Codility, CodeSignal. Proctored certification testing at scale: Mercer Mettl. A broad self-serve skills-test library on a small budget: TestGorilla. Pure sourcing and pipeline building: SeekOut, hireEZ. Whole-organisation skills inference: Eightfold. Internal talent marketplace: Gloat.

If the structured-assessment lane is the one you are in, you can book a demo or read the case studies first.

How to run this – a five-step sequence

1. Instrument the funnel before changing anything. Stage-by-stage pass-through and adverse impact ratio by group, with significance tests, under privilege where available. You cannot fix what you haven’t located, and most organisations find the loss at a stage they weren’t watching.
2. Separate the sourcing problem from the selection problem. Compare pool composition against labour-market availability. This single comparison determines which half of the remaining work is even relevant to you.
3. Fix the selection stage with structure, not sentiment. Job analysis, standardised questions, anchored rating scales, scores combined mechanically, overrides logged and reviewed, and hiring managers briefed on why the interview process is not theirs to improvise. Best evidence on both validity and adverse impact, lawful in every jurisdiction. Our guides to bias-free recruitment, eliminating recruitment bias and candidate evaluation go deeper on the mechanics.
4. Pressure-test the tooling. Run the vendor questions above – scope, subgroups, differential prediction, version control, training population, liability – and get the answers in writing.
5. Pilot on one role family and measure the funnel, not the headcount. One requisition family, one quarter, pre-agreed metrics: stage-level AIR, completion, offer-acceptance and 90-day retention by group. Decide on the data.

Conclusion

Three things are true at once. The problem is real and well measured: name-based penalties in the US, UK and Australia are large, concentrated in particular employers, and have not shrunk in fifty years of British data. The business case usually cited for fixing it does not survive replication, and it never needed to – a selection process that systematically misses capable people is a validity failure and a legal risk regardless of what it does to margin. And the interventions that work are the structural, unglamorous ones: structure your interviews, remove artificial barriers, instrument your funnel. If your diversity goals are currently written as headline representation numbers, rewrite them as stage-level ones. That single change will do more to improve diversity in your hiring outcomes than most of the programmes on the list above.

Which leaves an uncomfortable symmetry worth stating plainly. The practices most likely to get a US employer sued in 2026 are the demographic-outcome ones – slates, targets, restricted-eligibility programmes, tie-breakers. The practices with the best evidence behind them are the ones nobody is arguing about. If you are deciding where to spend the next twelve months, that is not a difficult choice.

If you want to see what a structured, measured version of the assessment stage looks like in practice, book a demo.

FAQs about diversity hiring

About Author

Barb Hyman

CEO & Founder