What is the four-fifths rule for AI hiring tools?

The four-fifths rule (also called the 80% rule) is an EEOC guideline for detecting adverse impact in selection procedures, including AI hiring tools. It compares the selection rate of a protected group against the group with the highest selection rate. If the ratio falls below 80%, the tool is presumed to have an adverse impact, which shifts the legal burden onto the employer to justify the tool as job-related and consistent with business necessity.

How do you calculate adverse impact for an AI screening tool?

Divide the selection rate of each demographic group by the selection rate of the group with the highest rate. For example, if an AI resume screener advances 60% of male applicants and 40% of female applicants, the ratio is 40/60 = 0.67, or 67% — below the 80% threshold, indicating adverse impact against female applicants that requires further justification.

Is passing the four-fifths rule enough to make an AI hiring tool legally safe?

No. The four-fifths rule is a practical screening threshold, not a legal safe harbor. Courts and the EEOC can find adverse impact through statistical significance testing even when the four-fifths ratio is technically met, especially with large applicant pools. It's a starting checkpoint, not a compliance finish line.

Who is responsible for adverse impact testing — the employer or the AI vendor?

Legally, the employer using the tool to make hiring decisions bears primary liability under Title VII, regardless of who built the software. Most enterprise AI hiring vendors now offer bias-audit reports, but relying solely on a vendor's audit without independently reviewing your own applicant-pool results is a common and costly compliance gap.

AI Hiring Tools and the Four-Fifths Rule 2026: Adverse Impact Testing Guide

Why the Four-Fifths Rule Matters More Now

The four-fifths rule has existed since the 1978 Uniform Guidelines on Employee Selection Procedures, long before AI hiring tools existed. But it's become the central legal battleground for algorithmic hiring because AI screening tools process thousands of applicants at once, creating exactly the kind of large, clean dataset that makes adverse-impact statistics easy to calculate — and easy to litigate.

A human recruiter who screens 40 resumes a week rarely generates a statistically meaningful adverse-impact case. An AI tool that screens 40,000 resumes a year across a company's entire hiring pipeline produces exactly the kind of dataset that turns a hunch into a lawsuit.

How to Calculate Adverse Impact: Step by Step

Calculate the selection rate for each group.

Selection rate = (number selected from group) ÷ (number of applicants from group). Do this separately for each demographic group you're comparing.

Identify the group with the highest selection rate.

This becomes your baseline — every other group's ratio is measured against it.

Divide each group's rate by the highest rate.

This produces the adverse impact ratio for each group, expressed as a percentage.

Flag any ratio below 80%.

A ratio under four-fifths (80%) of the highest-selection-rate group triggers the presumption of adverse impact.

Worked Example

An AI resume screener processes 500 applicants for a software engineering role: 300 men and 200 women. The tool advances 150 men (50% selection rate) and 70 women (35% selection rate).

Adverse impact ratio = 35% ÷ 50% = 0.70 = 70%

Since 70% is below the 80% threshold, this result presumes adverse impact against women applicants. The employer would need to show the screening criteria are job-related and consistent with business necessity — and that no less-discriminatory alternative exists that serves the same purpose.

Where the Four-Fifths Rule Falls Short

Employers who treat the four-fifths rule as a pass/fail compliance gate misunderstand its legal role. Courts have repeatedly found adverse impact even when a tool technically clears 80%, using statistical significance tests (like standard deviation analysis) that are more sensitive with large sample sizes — exactly the sample sizes AI hiring tools generate.

Small applicant pools can pass the four-fifths test by chance, masking real bias
Large applicant pools can fail statistical significance tests even above 80%
The rule doesn't account for intersectional impact (e.g., Black women specifically, not just women or Black applicants separately)
It measures outcomes at one stage only — bias can compound across multi-stage AI pipelines

Compliance Checklist for Employers Using AI Screening

Testing & Documentation

☐Run four-fifths analysis by race, sex, and age at minimum
☐Test intersectional subgroups, not just single categories
☐Re-run analysis quarterly or after any model update
☐Document business-necessity justification for the tool
☐Retain applicant-level data supporting your calculations

Vendor & Process Diligence

☐Request the vendor's independent bias-audit results
☐Confirm audit methodology tested your actual applicant pool
☐Identify a less-discriminatory alternative before deployment
☐Establish a human-review path for adverse-impact flags
☐Check applicable state audit laws (e.g., NYC Local Law 144)

Frequently Asked Questions

Does the four-fifths rule apply to every stage of an AI hiring pipeline?

It should be applied at each stage where the tool makes a selection decision — initial resume screening, interview scoring, and final ranking are each separate selection procedures that can independently produce adverse impact, and bias at each stage compounds through the funnel.

What sample size is needed for the four-fifths rule to be meaningful?

The EEOC's own guidance notes the rule is less reliable with small numbers — differences of one or two selections can swing the ratio dramatically in small pools. Most practitioners supplement four-fifths analysis with statistical significance testing once applicant pools exceed roughly 30 people per group.

Can an employer fix adverse impact by adjusting AI scores after the fact?

Adjusting scores based on protected-class membership after the model runs (sometimes called algorithmic affirmative action) creates its own legal risk under Title VII's prohibition on group-based preferences. The safer fix is addressing the underlying model or feature set causing the disparity, not post-hoc score adjustment.

Run the Numbers Before a Plaintiff Does

Adverse-impact analysis isn't optional diligence — it's the exact test opposing counsel will run against your hiring data in discovery. Running it proactively, quarterly, and across intersectional subgroups is the difference between catching a problem internally and explaining it in a deposition.

If your AI hiring vendor can't produce adverse-impact data broken out by demographic group for your specific applicant pool, that's a documentation gap worth closing now.