The Four Fifths Rule Is American. How to Check Hiring for Adverse Impact in the EU

If you use an applicant tracking system built in the US, you may have seen a dashboard turn a group's pass rate red below 0.8. For employers and staffing agencies hiring in the Netherlands, that is the wrong place to stop: EU law asks a different question, and answering it can need some of the most protected data you hold.

What is the four fifths rule, and where does it come from?

It is a US enforcement rule of thumb, set out in 29 CFR 1607.4(D). Under the Uniform Guidelines on Employee Selection Procedures (1978), a selection rate for any race, sex or ethnic group below four fifths (80%) of the rate for the group with the highest rate will generally be regarded by US federal enforcement agencies as evidence of adverse impact (eCFR, 29 CFR 1607.4). (US)

The four fifths rule formula:

The same paragraph qualifies the rule: smaller differences can count if significant in both statistical and practical terms, larger ones may not count if based on small numbers and not statistically significant, and numbers too small to be reliable may be judged over a longer period. The agencies' joint 1979 Questions and Answers call the 80% figure a rule of thumb, "not intended as a legal definition" but a practical way to focus enforcement on serious discrepancies (EEOC, Questions and Answers).

In the US, Executive Order 14281 of 23 April 2025 told federal agencies to deprioritise disparate impact enforcement (90 FR 17537); section 1607.4(D) itself was unchanged on eCFR on 21 October 2026. (US)

Does the four fifths rule apply in Europe?

No. EU and Dutch equal treatment law sets no numeric threshold.

EU law defines indirect discrimination in nearly identical terms across its main directives: an apparently neutral provision, criterion or practice that would put people with a protected characteristic at a particular disadvantage, unless objectively justified by a legitimate aim pursued by appropriate and necessary means. See Directive 2000/43/EC on racial or ethnic origin (Article 2(2)(b), applied to recruitment by Article 3(1)(a)), Directive 2000/78/EC on religion or belief, disability, age and sexual orientation (Article 2(2)(b)) and Directive 2006/54/EC on sex (Article 2(1)(b)).

The Dutch General Equal Treatment Act (AWGB) mirrors this: indirect distinction (indirect onderscheid) is a criterion that looks neutral but particularly affects people of, among other grounds, a certain religion, race, sex or nationality (Article 1), and it is banned in recruitment and job placement (Article 5) unless objectively justified (Article 2) (AWGB, version in force from 1 January 2026). The disability act (WGBH/CZ) uses a matching definition.

National rules may allow indirect discrimination to be shown by statistics (recital 15 of Directives 2000/43/EC and 2000/78/EC). Once a candidate shows facts from which discrimination may be presumed, the employer must prove there was no breach (Directive 2000/43/EC, Article 8; AWGB, Article 10). Statistics can start a case; a dated record of why each criterion is needed helps answer it.

How does the Court of Justice read statistics?

As evidence of disadvantage, judged case by case, never against a fixed threshold.

In Seymour-Smith and Perez (C-167/97, 1999), about a UK qualifying period of two years for unfair dismissal claims, the Court said the best approach is to compare the share of men and of women able to meet the requirement, and ask whether a considerably smaller percentage of women can. A lesser but persistent and relatively constant disparity over a long period could also count. The national court decides, including whether the statistics cover enough individuals, reflect more than chance or short term effects, and appear significant (paragraphs 59 to 62).

The figures were 77.4% of men and 68.9% of women, which on the face of it did not show a considerably smaller percentage (paragraphs 63 and 64). The four fifths ratio would be 0.89, above the US line, so the approaches agree here, but the Court used no ratio: it asked whether the gap was considerable, valid and lasting.

Justification has its own test. Under Bilka-Kaufhaus (C-170/84, 1986), a measure affecting far more women than men is justified only if it meets a real need of the employer and the means are appropriate and necessary. For hiring, that suggests a criterion that filters out one group more should meet a real need of the job, with no less disadvantaging way to check the same thing.

Our view: treat a gap as a question about the criterion. The four fifths ratio can serve as an internal trigger for a closer look, provided nobody mistakes it for the legal test.

What data may you use to compare groups?

Pass rates per step need no group data. Group comparisons often do.

Process data. Counts per step, criterion, channel, site and reviewer. You hold this already; analyse it in aggregate and limit access.

Protected grounds that are not special category data. Sex and age are protected in equal treatment law but are not GDPR special categories (GDPR, Article 9(1)). They still need a lawful basis, a defined purpose and minimisation.

Special category data. Data revealing racial or ethnic origin, religion, health or sexual orientation may be processed only under an Article 9(2) exception, and the obvious ones are hard to use in hiring:

Inferring a group from names or postcodes is unlikely to be a workaround: the Court of Justice has held that data revealing a special category indirectly, through deduction or cross referencing, are themselves special category data (C-184/20, 2022).

The Dutch Data Protection Authority (Autoriteit Persoonsgegevens, AP) reads the exceptions narrowly. In advice to the Dutch Tax Administration dated 14 August 2026, it found that without specific legislation no Article 9(2) exception covered the administration's bias testing. It accepted testing with data already held lawfully only as a temporary, strictly necessary measure with safeguards, said AI Act Article 4a would probably cover much of it, depending on whether the tools count as AI systems, and left the wider question open (AP advice, annex, published by the Senate on 1 September 2026). The advice concerns a public body, not an employer, but shows the regulator's stance.

What AI Act Article 4a allows, and what it does not

The Digital Omnibus on AI, Regulation (EU) 2026/1744, in force since 27 July 2026, inserted Article 4a into the AI Act. It lets providers and deployers of AI systems, high risk or not, exceptionally process special category data where strictly necessary to detect and correct bias. Conditions apply on top of the GDPR: other data, including synthetic or anonymised data, must not suffice; the data must be protected by measures including pseudonymisation and strict access controls, not passed to other parties, and deleted once the bias is corrected or the retention period ends, whichever comes first; and the records of processing must explain why it was strictly necessary.

Our view: this is our reading, not settled law. Article 4a concerns bias in relation to AI systems, so we do not read it as a general basis to collect ethnicity data to monitor a whole hiring process, including steps where no AI system is involved. Take legal advice before relying on it. The wider timetable is in our guide to the EU AI Act and recruitment.

Before collecting any group data, involve your data protection officer, counsel and works council, and consider a data protection impact assessment.

Which steps should you track?

Every step where a candidate can drop out or be turned away.

Pass rates per step show where candidates leave and which criterion removes them, without data on protected characteristics; any lawful group comparison later uses the same structure.

Illustrative example (invented numbers): one warehouse role, one quarter.

The knockout step removes 140 people, more than any other: find which question does most of that and whether the job needs it. Count withdrawals separately from rejections, and "evidence missing" separately from "requirement not met", since each needs a different fix. If a tool supports the review step, track how often reviewers overrule it; a rate near zero deserves a closer look (the tool should never be the one saying no). To map one role's steps, use our recruitment process audit, which takes an hour.

See how Radius handles human decisions, records and candidate data: read our trust page.

How do you compare two groups? A worked example

Calculate each group's pass rate, then the gap and its uncertainty.

Illustrative example (invented numbers), only where holding the group data is lawful for you. A knockout question asks for two years' warehouse experience. In group A, 14 of 40 applicants pass; in group B, 88 of 160.

The US rule would flag a ratio of 0.64. Under the EU approach, a gap of 20 points whose plausible range stays above zero is a reason to examine the criterion: is two years' experience needed, or would a short practical check show the same thing more directly (Three Years on Reach Trucks Tells You Less Than You Think)? The numbers are not proof of discrimination; they tell you where to look.

How many candidates do you need before a gap means anything?

There is no legal minimum, but small groups rarely tell you much.

Illustrative example (invented numbers): group A, 5 of 10 pass; group B, 8 of 12. The ratio of 0.75 would trigger a four fifths flag, yet Fisher's exact test gives p = 0.67, and a rough 95% interval for the gap runs from group A about 24 points better to about 58 points worse. The data fit both no difference and a large one.

Neither rescues small samples. In simulations, Collins and Morris (2008) found that all tests had low power at small sample sizes; Fisher's exact test was overly conservative, and the Z test for two proportions performed reasonably well except at extremely small sizes (Journal of Applied Psychology). (US research)

What do you do when a step shows a gap?

Review the criterion before anything else.

A different method is not automatically fairer. Work samples are often assumed to show small group differences, yet a 2008 meta analysis found a much larger gap between Black and white job applicants on overall work sample results than earlier estimates (a standardised difference of 0.73, against 0.38), though ratings of social skills showed small gaps (Roth and colleagues, 2008). (US research) What a check measures matters; check fairness again after every change. We make the same argument in No Hiring Tool Removes Bias. Including Ours.

What does this mean for staffing agencies and frontline employers?

High volumes make step tracking practical; client requirements need watching.

Pass rates show where candidates fall out, not why, and applicant pools differ by channel, site and season. Our guide to measuring quality of hire covers reading results over several hiring rounds.

Conclusion

The four fifths rule is useful arithmetic with a US legal history, not the European test. Track pass rates per step, compare groups only where the data is lawful and the numbers are large enough, and when a gap appears, fix the criterion rather than the statistics.

How Radius helps

Radius Hire is the AI hiring platform for frontline roles. It runs a short, structured voice interview in the language set for the role and shows recruiters the evidence behind each requirement.

Tracing a gap to a requirement needs the evidence behind it. For each requirement, Radius shows whether it is backed by a document or a check, rests only on what the candidate said, or is still missing (Match). Radius never rejects a candidate on its own: a recruiter reviews the evidence, can overrule anything Radius presents and makes every hiring decision. Each interview can be kept as a recording, a transcript and a summary, and each check carries its source and date, so a decision can be explained later (AI Interviewer). Candidate data is hosted in the EU and processed under a data processing agreement (trust).

This article does not describe a Radius feature for comparing groups of candidates. Whether and how you compare groups is your decision, made with your own legal advice.