Compliance · July 18, 2026 · 9 min read

Adverse Impact and the Four-Fifths Rule, Explained

Adverse impact and the four-fifths rule decide whether your hiring passes a bias audit. How the 80% ratio works, its failure modes, and how to monitor it.

← Part of Compliance-First Hiring AI: LL144 and the EU AI Act

You can run a hiring process that feels scrupulously fair and still be breaking the law, because in hiring, fairness is judged by outcomes, not intentions. If you screen, assess, or rank candidates, adverse impact and the four-fifths rule are the concepts your bias audit turns on: they are how a neutral-looking process gets tested for discriminatory effect. This guide is for talent leaders, recruiters, and the HR or legal owners of hiring-tool compliance who need to understand what the ratio measures, where it misleads, and how to monitor it without becoming statisticians.

Why the stakes are high: adverse impact is measured on results, so a tool everyone believed was neutral can still screen out a protected group at a disproportionate rate and expose the employer. Regulation like NYC's Local Law 144 and the EU AI Act has made that exposure concrete, with bias audits and impact analysis moving from good practice to legal requirement. This is a deep dive from our compliance-first hiring pillar: the concept every audit references and few teams measure well.

What is adverse impact?

Adverse impact, sometimes called disparate impact, occurs when a neutral-looking selection process screens out a protected group at a substantially higher rate than others. The distinguishing feature is that it can happen with no intent to discriminate. A knockout question, a timed test, a résumé filter, or an automated score can all produce lopsided outcomes even when nobody designed them to. That is precisely why adverse impact has to be measured rather than assumed away: good intentions are not evidence, and the people running the process are rarely the ones who notice its effect.

This matters at every stage, not just the final decision. Each step in your funnel — sourcing, résumé screening, assessments, interviews, offers — is its own selection event with its own pass rates. A process can look balanced overall and still contain one stage that quietly filters out a protected group. Analysing the whole funnel as a single number can hide the very stage causing the problem, which is why practitioners look stage by stage.

It is worth being precise about the legal concept behind the term. Adverse impact is the practical, measurable side of what US law calls disparate impact: a facially neutral practice that falls more harshly on a protected group and cannot be justified by business necessity. Intent is not required for liability. That is the crucial difference from disparate treatment, where the question is whether someone was deliberately treated worse. Because intent drops out of the picture, an employer cannot defend an adverse-impact finding by pointing to a sincere belief that the process was fair — only by showing the practice is job-related and consistent with business necessity, and that no equally valid, less discriminatory alternative was available.

How does the four-fifths rule work?

The four-fifths rule is the most widely used heuristic for spotting adverse impact. The method is deliberately simple. Work out each group's selection rate, identify the group with the highest rate, then compare every other group against it. If any group's rate falls below 80% — four-fifths — of the highest group's rate, that is a flag worth investigating.

A worked example

Suppose the highest-selecting group passes an assessment stage at 60%, and another group passes at 40%. Divide 40 by 60 and you get 0.67, comfortably below the 0.8 threshold. That gap is a signal that the stage may be producing adverse impact and deserves a closer look. It does not, on its own, prove the assessment is discriminatory or invalid; it tells you where to point your attention.

The rule traces back to the 1978 US Uniform Guidelines on Employee Selection Procedures, issued jointly by federal enforcement agencies. It was designed as a practical screen that non-statisticians could apply, not as a definitive legal test. That origin explains both its usefulness and its limits: it is blunt on purpose, and it was never meant to be the last word.

Two practical points follow from how the calculation is set up. First, you always compare against the highest-selecting group, not against an average or an overall rate, so the reference point can shift between hiring rounds as different groups come out on top. Second, the rule is applied per selection stage and per relevant characteristic, which is why a single company-wide fairness number tells you very little. The value of the four-fifths rule is that it is cheap enough to run at every stage, often — and that frequency, not the precision of any single figure, is where its usefulness lies.

Where the four-fifths rule breaks down

Because the rule is a heuristic, it has well-known failure modes, and a compliance-first team needs to know all of them before leaning on the number.

• Small numbers. With few candidates in a group, a single hire or rejection can swing the ratio dramatically. A four-in-five versus three-in-four outcome can breach the threshold on tiny samples where the difference is statistically meaningless.
• False comfort. A process can clear 0.8 and still be unfair, or contain a discriminatory stage masked by a balanced overall figure. Passing the ratio is not a clean bill of health.
• False alarms. A ratio can fall below 0.8 for reasons unrelated to the assessment — pipeline composition, self-selection, or upstream sourcing — so a breach is a prompt to investigate causes, not to condemn a tool.
• Statistical significance. Regulators and courts increasingly look past the raw ratio to significance testing, asking whether an observed gap is likely to be real rather than noise. The four-fifths figure is a starting point, not the whole analysis.

This is exactly why regulators, and audit regimes like NYC's Local Law 144, treat a failed ratio as a trigger to investigate rather than an automatic finding of discrimination. When a stage flags, the question becomes whether the selection procedure is job-related and consistent with business necessity — the legal standard that a demonstrably relevant, validated assessment is built to meet. A generic test scraped from a shared bank is far harder to defend on that ground than one tied to the actual work of the role.

How to monitor adverse impact continuously

Impact ratios are not static. They drift as your candidate pool shifts, as you open new roles, and as you tune each stage, so a one-time check is never enough. The teams that stay ahead of this treat it as ongoing instrumentation rather than an annual scramble the week before an audit is due. Three habits make the difference.

Measure per stage, not just at the end

Track selection rates at each stage — screen, assessment, interview, offer — because adverse impact usually enters at one specific step. An end-to-end ratio can look acceptable while a single stage is doing the damage. Stage-level monitoring tells you not only that something is wrong but where, which turns a vague fairness concern into a fixable engineering problem.

Keep the records an audit would ask for

Fairness you cannot evidence is fairness you cannot defend. Retain the selection data, the reasoning behind decisions, and the validation basis for each assessment stage. When a candidate or regulator asks why someone was screened out, a defensible answer rests on documented job-relatedness, not on an assurance that the team meant well.

Keep a human accountable

Automated stages should feed decisions a person can review, override, and explain. Meaningful human review is both good practice and, increasingly, a legal requirement — and it depends on the process producing something reviewable. This is where assessment design and compliance meet: build stages that generate concrete, job-related evidence, and the human in your loop has something honest to work with.

Set a cadence, not a calendar reminder

How often is often enough depends on volume, not on the calendar. A high-throughput role that hires monthly needs monitoring far more frequently than one you fill twice a year. The practical rule is to review whenever you have accumulated enough decisions for a stage's ratio to mean something, and to treat the annual audit as a formalisation of monitoring you are already doing, not as the only time anyone looks. A number you compute once a year is a number that can be wrong for eleven months.

Small samples deserve particular care in this cadence. When a group has too few candidates for the four-fifths ratio to be stable, resist the urge to either dismiss a breach as noise or treat it as proof. The honest response is to note it, watch whether it persists as numbers grow, and combine the raw ratio with a sense of statistical significance before drawing conclusions. Over-reacting to small-sample swings burns credibility; ignoring a persistent pattern because 'the numbers are small' is how real problems hide.

How adverse impact fits your wider bias strategy

Measuring adverse impact is the audit half of fair hiring; the design half is giving bias fewer places to enter in the first place. The two reinforce each other. Structured, standardised stages produce cleaner data to monitor, and monitoring tells you which structured stages still need work. Our guide to reducing bias in hiring covers the design side stage by stage — inclusive job ads, skills-based screening, and behaviourally-anchored scoring — while this article covers the measurement that proves those changes are working.

The pattern across modern hiring regulation, from NYC to the EU, is consistent: tell candidates when AI is involved, keep a human meaningfully in the loop, measure outcomes for adverse impact, and keep records. Teams that build to that common core once, treating compliance as a design constraint rather than a paperwork layer, spend far less than teams patching each requirement after the fact.

Where H-Evaluate fits

H-Evaluate was built compliance-first for exactly this problem. Assessments are generated per job description through a quality-gated process, so every question is job-related by construction — the foundation of any job-relatedness defence when a stage flags under the four-fifths rule. AI Sandbox work samples produce reviewable artefacts rather than opaque scores, so the human in your loop reviews the actual work a candidate did. Selection data is captured per role and per stage, which is what continuous adverse-impact monitoring assumes you have.

None of that removes your obligation to run the analysis and confirm requirements with counsel — no tool can. What an AI-native platform can do is make the evidence a standing by-product of how you hire, rather than something you reconstruct under deadline. If you are rethinking how you measure fairness, that is the place to start.

The four-fifths rule will not tell you your hiring is fair. It will only tell you where to look — and whether you built a process you can explain to a candidate, an auditor, or a court.