The Four-Fifths Rule: How to Check Your Hiring Process for Adverse Impact
8 min read · Last reviewed 1 Aug 2026
General information, not legal advice. Laws in this area change; verify against the official sources at the end of this guide and confirm specifics with employment counsel.
The four-fifths rule is the closest thing US hiring law has to a smoke detector: a simple ratio that flags when a screening step selects one demographic group at a much lower rate than another. It comes from the 1978 Uniform Guidelines on Employee Selection Procedures and has outlived every wave of hiring technology since, because it does not care how the selection happened, only what came out. A biased keyword filter, a biased AI model, and a biased manager all trip the same wire.
If you use any automated screening, the four-fifths math is worth knowing for three reasons: it is the standard courts and agencies reach for first in disparate-impact analysis, it is the core computation inside an NYC Local Law 144 bias audit, and it takes about fifteen minutes to run on your own funnel. This guide walks the math, what a flag means, and where federal enforcement stands in 2026. General information, not legal advice.
What the rule says
The Uniform Guidelines' formulation: a selection rate for any race, sex, or ethnic group that is less than four-fifths (80%) of the rate for the group with the highest rate will generally be regarded as evidence of adverse impact.
Unpack the terms. A selection rate is simply the fraction of a group's applicants who pass the step you are measuring: made the shortlist, got the interview, received the offer. The comparison is between each group's rate and the highest group's rate, expressed as a ratio called the impact ratio. Below 0.80, the guideline treats the disparity as worth explaining.
Two scoping notes. First, the rule applies per step, not just to final hires: a screening stage can create adverse impact even if the offers end up balanced, and step-by-step analysis is exactly how an auditor will look at an AI screening tool. Second, the rule is a rule of thumb, not a statute: it is the evidentiary threshold agencies and courts historically used to decide when a disparity deserves scrutiny, and small samples can trip it by chance.
The math, with a worked example
Suppose a job posting draws 150 applicants: 100 men and 50 women. Your screening step (human, AI, or both) advances 40 men and 12 women.
Selection rate for men: 40 / 100 = 40%. Selection rate for women: 12 / 50 = 24%. The highest rate is the men's 40%, so the impact ratio for women is 24 / 40 = 0.60. That is well under 0.80: the four-fifths flag fires, and if this pattern held up statistically, an agency or plaintiff would treat the screening step as showing evidence of adverse impact against women.
How many more selections would clear the flag? The threshold rate is 0.8 x 40% = 32%, which on 50 applicants means 16 selections. The gap between 12 and 16 women is the size of the problem.
Run the same computation for race/ethnicity categories, and, if you want the fuller LL144-style picture, for intersectional groups (for example, Black women as a category rather than only Black candidates and only women). Watch the denominators: with 7 applicants in a group, one person flips the ratio wildly, which is why the Uniform Guidelines and every serious auditor pair the ratio with a sample-size caveat and, at larger scale, statistical significance tests.
What a flag means, and what it does not
A ratio under 0.80 is not a finding of illegal discrimination. It is evidence that shifts the burden: in a disparate-impact framework, the employer then has to show the selection procedure is job-related and consistent with business necessity, and the challenger can respond that a less discriminatory alternative existed.
That structure tells you exactly how to respond to a flag. First, check the data: small samples, misclassified outcomes, and one-off hiring bursts create false alarms. Second, find the mechanism: which criterion is driving the disparity? A requirement that is genuinely necessary for the job (a license, a certification) can produce a disparity you can defend; a proxy criterion (an arbitrary degree requirement, an address-based filter, a writing-style preference that penalizes non-native speakers) is where the danger lives. Third, fix or justify: drop or reweight criteria you cannot defend, document the business necessity of the ones you keep, and re-run the numbers.
A flag you found yourself, investigated, and fixed is a good story. A flag a plaintiff's expert found first, in data you never looked at, is a bad one. That asymmetry is the entire argument for self-checking.
Where federal enforcement stands in 2026
The federal picture shifted in 2025: the EEOC removed its AI-specific technical guidance from circulation, and an executive order directed federal agencies to deprioritize disparate-impact enforcement. It would be a mistake to read that as the four-fifths rule no longer mattering, for three reasons.
First, Title VII itself did not change: disparate-impact liability remains in the statute, and private plaintiffs can and do sue without the EEOC's help. The Mobley v. Workday litigation, which advanced to a collective action covering age-discrimination claims against a screening vendor, is the live demonstration that algorithmic screening claims have a path through the courts regardless of enforcement posture in Washington.
Second, the states moved the other way. NYC Local Law 144 made impact ratios a published, annual, audited requirement. California made anti-bias testing evidence in FEHA claims. Illinois attached effect-based liability to AI hiring. The four-fifths computation is embedded in all of them.
Third, enforcement postures are temporary and hiring records are not: screening decisions you make today are discoverable for years, under whatever posture exists later. The math is cheap; run it.
How to run the check on your own funnel
You need two ingredients: outcomes per candidate (who advanced past the step) and demographic categories per candidate. The outcomes you have. The demographics must come from voluntary self-identification collected separately from the application, the way EEO surveys already work; never infer race or sex from names, photos, or writing style, which is itself a discrimination and accuracy risk.
Mechanically: export your screening results as a table of candidate, outcome; join your self-ID data; compute each group's selection rate; divide by the highest rate; investigate anything under 0.80 with a decent sample behind it. Repeat per hiring stage and per role family a few times a year and whenever you change tools or criteria.
SiftFirst gives you the pieces for free. The audit-ready export produces the per-candidate scores and shortlist outcomes with no demographics in it. The bias audit self-check at /tools/bias-audit-check then runs the four-fifths math entirely in your browser: you add the demographic columns yourself and paste the CSV, nothing is uploaded or stored, and you get impact ratios per group with the under-0.80 rows flagged. It is a self-check, not an official audit (NYC requires an independent auditor for that), but it is the same math, and it turns "we never looked" into "we check quarterly."
Key takeaways
- ✓The four-fifths rule flags adverse impact when any group's selection rate is under 80% of the highest group's rate; it applies per hiring step, not just to final offers.
- ✓The math is three divisions: group selection rate, divided by highest group's rate, compared to 0.80. Small samples flip ratios easily, so pair it with sample-size sanity checks.
- ✓A flag is not a finding of discrimination; it shifts the burden to showing the criterion is job-related, which is defensible for real requirements and dangerous for proxies.
- ✓Federal enforcement posture softened in 2025, but Title VII private claims, Mobley v. Workday, and the NYC/California/Illinois laws all keep the computation legally live.
- ✓Demographics must come from voluntary self-ID collected separately, never inferred; a browser-based self-check on your own funnel takes about fifteen minutes.
Screening built for these rules
SiftFirst scores candidates against criteria you set, quotes the resume line behind every score, never auto-rejects, and exports the records these laws expect. The candidate notice generator and bias audit self-check are free.
FAQ
Is passing the four-fifths test a safe harbor?
No. A ratio above 0.80 does not immunize a process (statistically significant disparities can matter even above the line, and other evidence can support a claim), and a ratio below it does not condemn one (job-related necessity is a defense). Treat it as a well-established screening threshold: cheap to compute, informative, and the first thing an auditor or expert will calculate.
Where do I get the demographic data to run the check?
From voluntary self-identification, collected separately from the application materials, the same way EEO-1 style surveys work. Store it away from scores, and never let reviewers see it during screening. Do not infer demographics from names, photos, schools, or writing style: inference is inaccurate and is itself a discrimination risk. If you have no self-ID data, you can still prepare the outcome export and start collecting self-ID going forward.
Does the four-fifths rule apply to small businesses?
Title VII covers employers with 15 or more employees, but state equivalents reach lower (California FEHA at 5, Illinois at 1), and NYC LL144 has no size threshold for its audit requirement. Below every threshold, the math still tells you whether your screening is skewed, which you want to know before it compounds. The check costs minutes; run it regardless of headcount.