Disparate Impact in ML: The Four-Fifths Rule Explained

Contents

Figure 1: Disparate Impact in ML: The Four-Fifths Rule Explained — one division, one threshold, and a legal test that decides whether your model's outcomes get investigated

A hiring model never sees a candidate's race or sex, passes every code review, and still selects one group at far lower rates than another. Under U.S. anti-discrimination law, that pattern can matter even though nobody intended it.

The doctrine is called disparate impact, and the most widely used screening test for it is the four-fifths rule. This guide explains what the rule says, how to compute it, why it is a blunt instrument, how it is used in ML audits, and where the law stands in late 2026. One caveat before anything else: I am not a lawyer, this is general information rather than legal advice, and the legal situation here is unusually unsettled, so involve counsel for any real decision.

Disparate Treatment vs Disparate Impact

Disparate treatment is intentional discrimination: treating someone differently because of a protected characteristic. Disparate impact, also called adverse impact, is different. A facially neutral practice, such as a test, a screening algorithm or a credit score cutoff, disproportionately excludes members of a protected group, and the employer cannot justify it.

The U.S. Supreme Court recognised this theory under Title VII in Griggs v. Duke Power Co. in 1971, and the Civil Rights Act of 1991 codified a burden-shifting framework:

1. The plaintiff shows that a practice causes a disparity in outcomes for a protected group.
2. The employer must show the practice is job-related and consistent with business necessity.
3. Even then, the plaintiff can prevail by showing a less discriminatory alternative that would serve the employer's legitimate needs.

That third step matters a great deal for ML, because models can be rebuilt in many ways, as the alternatives section below covers. A screening tool is not a fixed object with one defensible form.

The Four-Fifths Rule

The rule comes from federal guidelines on employee selection procedures, codified at 29 CFR 1607.4(D). It says that a selection rate for any group that is less than four-fifths (80%) of the rate for the group with the highest rate is generally regarded as evidence of adverse impact.

The calculation takes three steps:

1. Selection rate for each group = number selected divided by number of applicants.
2. Impact ratio for each group = that group's selection rate divided by the highest group's selection rate.
3. Flag any impact ratio below 0.80.

A Worked Example

Suppose 200 applicants from Group A and 100 from Group B go through an automated screen:

Group	Applicants	Selected	Selection rate	Impact ratio	Flag
A	200	120	60%	1.00 (reference)	—
B	100	36	36%	0.36 / 0.60 = 0.60	Below 0.80, adverse impact
B (alternative)	100	54	54%	0.54 / 0.60 = 0.90	Passes the screen

Group B needs 48 selections to reach exactly 0.80, so the failing case is two hires short of the line. That arithmetic is worth internalising, because it explains why the rule is so sensitive to small samples.

What the Rule Is Not

Two points about what the rule is:

• It is a rule of thumb for flagging, not a legal conclusion in either direction.
• Passing it is not a safe harbour, since courts and agencies can still find adverse impact in a disparity that clears 80%, particularly when it is statistically significant.
• Failing it does not automatically mean illegality, because the employer may be able to show business necessity.

Where the Rule Gets Shaky

Weakness	What goes wrong
Small samples	With few applicants, one extra hire swings the ratio dramatically, and a ratio above 0.8 can still be statistically significant in a large sample
Reference group dependence	The comparison group is whichever group has the highest rate, so a tiny group with a lucky rate can make everyone else look bad
Base-rate sensitivity	The ratio behaves differently when the screen is lenient than when it is strict, which the simulation below shows
Ignores qualifications	It compares raw selection rates, not whether candidates were equally qualified, which is why the legal framework adds job-relatedness and alternatives analyses
One group at a time	Intersectional groups, such as women over 50, can be disadvantaged even when each single attribute looks fine
One decision point	Hiring is a funnel of stages, and a disparity can arise at any of them

The guidelines themselves note that smaller differences may be significant with large samples, and larger ones may not be with small samples, which is why practitioners pair the ratio with significance tests rather than treating it as a verdict.

Where It Applies to ML

The classic setting is employment, and ML makes it pressing: résumé screeners, video interview scoring, assessment tools and ranking systems all produce selection decisions, at a volume where a small gap becomes thousands of people.

Other U.S. regimes, including lending under the Equal Credit Opportunity Act and housing under the Fair Housing Act, have their own disparate-impact doctrines, and the federal enforcement posture in several of them has also shifted recently, so check current status for your domain. Outside the U.S., concepts like indirect discrimination in the UK Equality Act and the EU AI Act's bias-examination requirements play similar roles but with different tests.

The Legal Landscape in Late 2026

This is the part most likely to change, so verify each point before you rely on it:

Instrument	What it says	Status to verify
Title VII and 29 CFR 1607	Disparate-impact theory and the four-fifths rule	Unchanged on the books as of the sources I found
April 23, 2025 executive order, "Restoring Equality of Opportunity and Meritocracy"	Directs agencies to move away from disparate-impact enforcement	Changes enforcement priorities, not the statute; private plaintiffs retain standing
Reported June 2026 OLC opinion	Reportedly calls Title VII disparate-impact enforcement unconstitutional	Single secondary source; such opinions bind the executive branch, not courts, so verify
New York City Local Law 144	Annual independent bias audits of automated employment decision tools using impact ratios, plus public summaries and candidate notice	A December 2025 state comptroller audit raised enforcement concerns
California FEHA automated-decision-system regulations	Disparate-impact framework, liability for vendors acting as employer agents, four-year record retention, anti-bias testing relevant to a defence	Effective 1 October 2025
Illinois HB 3773	Prohibits AI use that has the effect of discriminating, bans zip codes as proxies, requires notice	Effective 1 January 2026
Texas 2026 law	Requires intent	Narrower than disparate impact by design
Colorado AI Act	Delayed, scaled back and, according to one report, repealed and re-enacted in May 2026	Sources give different effective dates, so read the current text

The practical upshot: even if federal enforcement recedes, private lawsuits, state attorneys general and local laws can still reach algorithmic disparities, and the four-fifths rule remains the lingua franca for audits. A compliance strategy that watches only Washington will miss the rulebooks that actually apply to it.

Measuring It in Python

First, the core table: selection rates, impact ratios and flags, plus significance and an interval, because the ratio alone is not enough.

import numpy as np
import pandas as pd
from scipy.stats import fisher_exact

def impact_table(df, group_col, selected_col):
    t = (df.groupby(group_col)[selected_col]
           .agg(applicants="size", selected="sum")
           .assign(rate=lambda d: d.selected / d.applicants))
    t["impact_ratio"] = t["rate"] / t["rate"].max()
    t["flag_4_5"] = t["impact_ratio"] < 0.8
    return t.sort_values("impact_ratio")

def fisher_vs_reference(t, reference=None):
    """Significance of each group's disparity relative to the highest-rate group."""
    ref = reference or t["rate"].idxmax()
    p_values = {}
    for g in t.index.drop(ref):
        table = [[t.loc[g, "selected"], t.loc[g, "applicants"] - t.loc[g, "selected"]],
                 [t.loc[ref, "selected"], t.loc[ref, "applicants"] - t.loc[ref, "selected"]]]
        p_values[g] = fisher_exact(table)[1]
    return pd.Series(p_values, name="fisher_p")

def ratio_ci(df, group_col, selected_col, group, ref, iters=2000, seed=0):
    """Bootstrap interval for the impact ratio of one group against the reference."""
    rng = np.random.default_rng(seed)
    a = df.loc[df[group_col] == group, selected_col].to_numpy()
    b = df.loc[df[group_col] == ref, selected_col].to_numpy()
    ratios = []
    for _ in range(iters):
        rb = rng.choice(b, size=len(b)).mean()
        if rb > 0:
            ratios.append(rng.choice(a, size=len(a)).mean() / rb)
    return np.percentile(ratios, [2.5, 97.5])

Always report the ratio with an interval or a p-value, since a ratio of 0.78 from 40 applicants and 0.78 from 40,000 applicants mean very different things. The bootstrap interval is the part reviewers forget, and it is the part that tells you whether a flag is a finding or a rounding artefact.

See the Base-Rate Sensitivity

This simulation gives Group B a score distribution shifted slightly lower, then applies progressively stricter cutoffs:

rng = np.random.default_rng(0)
n_a, n_b = 6000, 2000
df = pd.DataFrame({
    "group": ["A"] * n_a + ["B"] * n_b,
    "score": np.concatenate([rng.normal(0.00, 1, n_a), rng.normal(-0.25, 1, n_b)]),
})

for top_frac in (0.5, 0.2, 0.05):
    cutoff = np.quantile(df["score"], 1 - top_frac)
    df["selected"] = (df["score"] >= cutoff).astype(int)
    t = impact_table(df, "group", "selected")
    print(f"select top {top_frac:>4.0%}: impact ratio of B = {t.loc['B', 'impact_ratio']:.2f}")

The same underlying score gap produces a mild disparity under a lenient screen and a severe one under a strict screen. This is why a model that passes at one operating threshold can fail at another, and why you should audit at the thresholds you actually use and across the range you might use later. A screening tool sold with a "top 10%" default will look different from the same tool at "top 50%", even though the model is identical.

Using Libraries

Fairlearn's demographic_parity_ratio computes the lowest group selection rate divided by the highest, which is the same quantity as the minimum impact ratio, and MetricFrame slices any metric by group. IBM's AIF360 provides a disparate-impact metric on its dataset-metric classes. Check each library's current documentation, and verify their numbers against the transparent calculation above — a library is a convenience, not an authority, and version drift is real.

Getting Protected-Attribute Data

You cannot measure group disparities without group labels, which creates a tension: you want to audit by race and sex without using them to make decisions.

1. Collect demographic data voluntarily through self-identification.
2. Store it separately from the decision pipeline.
3. Use it only for auditing, with access controls and retention limits.
4. Include intersectional categories where the data allows.

Inferring demographics from names and addresses, a technique used in some lending contexts, introduces its own error and legal questions, so treat inference as a last resort and disclose its limits. Small groups need special handling too, because very few members make the ratio unstable; New York City's rules, for example, allow excluding categories that make up a tiny share of the data from the calculation while requiring disclosure, though check the current rules.

When the Ratio Falls Below 0.8

Do not panic, and do not reach for a quick fix:

1. Check the data and the calculation: correct groups, correct denominator (all applicants, not only those who reached a later stage), and enough samples.
2. Check significance and stability, with intervals and over time.
3. Locate the disparity in the funnel. Which stage produces it?
4. Investigate the drivers: which features and proxies (zip codes, schools, gaps in employment, writing style) contribute? Attribution and partial dependence tools help here.
5. Assess validity. Is the model's output actually related to job performance, with evidence? Under the legal framework, the burden of justifying a disparity rests on job-relatedness and business necessity.
6. Search for less discriminatory alternatives. Since the third legal step turns on whether an alternative exists, treat this as an engineering task: retrain with different features, constraints or model families, compare validity and adverse impact across candidates, and document the search. Different models with similar accuracy can have very different disparities, a phenomenon sometimes called model multiplicity.
7. Consider mitigation methods — pre-processing such as reweighing, in-processing with fairness constraints, or post-processing. Be careful with the last: group-specific cutoffs or score adjustments can themselves create legal exposure in employment, since Title VII restricts adjusting scores or using different cutoffs on the basis of protected characteristics, so get legal advice before any group-aware correction. The pre-, in- and post-processing comparison shows what each family buys and costs.
8. Document everything: the analysis, alternatives considered and rationale. Documentation supports a defence and your own governance, and the AI risk assessment template is a reasonable place to record it.

Beyond the Four-Fifths Rule

The ratio is a screen, and a complete audit adds:

• Statistical significance tests, such as Fisher's exact test or a z-test for proportions.
• Practical significance: effect sizes and the number of people affected.
• Scoring-rate analysis for tools that output scores rather than decisions. New York City's rules, for instance, define a scoring-rate variant based on the share of candidates scoring above the median.
• Error-rate fairness metrics: equal opportunity (similar true positive rates among qualified candidates) and equalised odds, which speak to whether qualified people are treated alike, something raw selection rates cannot show.
• Calibration across groups.

These criteria cannot all be satisfied at once when base rates differ, a result covered in the fairness trade-offs literature, so decide which matter for your context and be transparent about the choice. The bias case studies show what happens when a metric that looked fine hid the harm.

Common Pitfalls

1. Treating 0.8 as a safe harbour, when passing does not prevent liability and failing does not prove it.
2. Auditing once at launch and never again. Input distributions, thresholds and models all change.
3. Ignoring sample size, and reporting ratios with no uncertainty.
4. Believing that removing the protected attribute removes the problem. Proxies like zip code, school and word choice carry the same information.
5. Applying group-specific thresholds without legal review.
6. Auditing only training labels and not actual outcomes.
7. Looking only at the final decision, not at each stage of the funnel.
8. Skipping intersectional analysis, so a subgroup's harm never appears in any single-attribute table.
9. Assuming a vendor's tool is the vendor's problem, when state rules can make you responsible for it.
10. Confusing disparate treatment with disparate impact, and so addressing neither properly.

Short checklist:
[ ] decision points and affected groups identified
[ ] demographic data collected voluntarily and stored separately
[ ] selection rates, impact ratios, intervals and significance tests computed at real thresholds
[ ] funnel stages and intersections examined
[ ] drivers and proxies investigated, with validity evidence for job-relatedness
[ ] less discriminatory alternatives searched and documented
[ ] legal review before any group-aware mitigation
[ ] monitoring, retraining triggers, vendor audit rights and record retention in place

Check the edition date before you buy, and pair any book on the law with the primary sources above, because the enforcement picture moves faster than print.

Unlock AI That Actually Works

Get lifetime access to GPT-6 Astra, Claude Fable 5.1, Gemini 3.5, Grok 4.5, and more — all in one platform. Build websites, apps, videos, content, and digital products from a single command. No monthly fees. No tool-hopping.

Click here to get GPTAstra Max now — one-time payment, lifetime access.

Frequently Asked Questions

What is the four-fifths rule?

It is a screening test from the federal Uniform Guidelines on Employee Selection Procedures, codified at 29 CFR 1607.4(D). It says a selection rate for any group that is less than four-fifths, or 80%, of the rate for the group with the highest rate is generally regarded as evidence of adverse impact. You compute each group's selection rate, divide it by the highest group's rate, and flag anything below 0.80. It is a rule of thumb for flagging, not a legal conclusion, and it does not account for qualifications, sample size or the choice of reference group.

Does passing the four-fifths rule mean my model is legal?

No. Passing is not a safe harbour, because courts and agencies can still find adverse impact in a disparity that clears 80%, particularly when the sample is large enough for the difference to be statistically significant. Failing is not proof of illegality either, because the employer may show the practice is job-related and consistent with business necessity. The rule is a first screen that tells you where to look, and the legal analysis around it involves qualification evidence, alternatives and jurisdiction-specific rules.

How do you calculate adverse impact in Python?

Group by the protected attribute, count applicants and selections, divide selected by applicants to get each group's selection rate, then divide each rate by the maximum rate to get the impact ratio, and flag ratios below 0.8. Always attach an interval or a p-value: a ratio of 0.78 from 40 applicants and one from 40,000 mean very different things, and Fisher's exact test against the highest-rate group is a reasonable default. Libraries such as Fairlearn and AIF360 compute the same quantity, but verify their output against the transparent calculation.

Is disparate impact still enforced in 2026?

Not reliably at the federal level, and clearly yes at the state and local level. Title VII and the four-fifths regulation remain on the books, but an April 2025 executive order directed agencies to deprioritise disparate-impact enforcement, and the EEOC's AI-related technical assistance has been removed from its site. One June 2026 report describes a Justice Department Office of Legal Counsel opinion calling the theory unconstitutional, which comes from a single secondary source and binds the executive branch rather than courts. Meanwhile New York City, California and Illinois impose their own audit and notice duties.

Can I audit for bias without collecting race and sex data?

You cannot measure group disparities without group labels, so the usual compromise is to collect demographic data voluntarily through self-identification, store it separately from the decision pipeline, and use it only for auditing. Inferring demographics from names and addresses introduces its own error, plus legal questions, so treat it as a last resort and disclose its limits. Small groups need special handling because their ratios are unstable, and laws such as New York City's allow excluding tiny categories from the calculation while requiring disclosure.

Does removing protected attributes fix disparate impact?

No, and this is one of the most common mistakes. A hiring model that never sees race or sex can still select one group at far lower rates, because proxies such as zip code, school, employment gaps and writing style carry the same information. Disparate impact is defined by outcomes, not inputs, so the only way to know is to measure selection rates by group at the thresholds you actually use. That is also why the doctrine can reach a facially neutral practice at all.

Sources

• 29 CFR 1607.4, Uniform Guidelines on Employee Selection Procedures — the four-fifths text.
• Griggs v. Duke Power Co., 401 U.S. 424 (1971) — the case that recognised disparate impact under Title VII.
• Executive Order, Restoring Equality of Opportunity and Meritocracy — the April 2025 order on enforcement priorities.
• Fairlearn: demographic_parity_ratio — the ratio computation described above.
• AI Fairness 360 documentation — datasets, metrics and mitigation classes.
• NYC Department of Consumer and Worker Protection: automated employment decision tools — Local Law 144 guidance. This page blocks bot requests, so open it in a browser.

Links checked on 21 October 2026. California's FEHA automated-decision-system regulations and Illinois HB 3773 are cited from secondary reporting; read the primary text on calcivilrights.ca.gov and ilga.gov before relying on them.

Wrapping Up

Disparate impact asks whether a neutral-looking practice disproportionately excludes a protected group, and the four-fifths rule is the standard first screen: divide each group's selection rate by the highest group's rate and look hard at anything under 0.80. It is easy to compute and easy to misuse, since it is sensitive to sample size, base rates and the choice of reference group, and it says nothing about qualifications. In ML, the useful response to a flag is an investigation: significance, funnel location, drivers, validity, and a documented search for less discriminatory alternatives.

Is the legal risk going away because federal enforcement priorities changed? Not reliably. The statute remains, private plaintiffs retain their claims, and state and local rules keep expanding. Compute your impact ratios at your real thresholds, put intervals on them, and talk to counsel about what to do with what you find.

What is your model's impact ratio at the threshold you actually ship, and would it survive being computed per funnel stage?

Related Articles

• Bias Mitigation Techniques: Pre-, In-, and Post-Processing Compared
• Algorithmic Bias Examples in Hiring, Lending and Healthcare
• AI Ethics Principles: Fairness, Accountability and Transparency
• How to Run an AI Risk Assessment: Template and Process
• Interpretability vs Explainability in Machine Learning: What's the Difference?