The four-fifths rule, also called the 80% rule, is a screening heuristic that flags possible adverse impact when the selection rate for one group is less than 80 percent of the rate for the most selected group. Originating in United States employment law, it has become a standard first check in AI ethics and algorithmic fairness work for hiring, lending and other automated decisions.
The rule is a rule of thumb, not a legal verdict. A ratio below 0.8 signals that a selection procedure deserves closer scrutiny, while a ratio above 0.8 does not prove that the procedure is fair.
How the four-fifths rule is calculated
The four-fifths rule compares selection rates across groups defined by a protected attribute such as sex, race or age. The selection rate of a group is the number of people selected divided by the number of applicants from that group.
The analyst identifies the group with the highest selection rate and divides every other group’s rate by it. The result is the impact ratio. If any group’s impact ratio falls below 0.80, the procedure shows evidence of adverse impact under the rule. In machine learning terms, the impact ratio is the ratio form of demographic parity, applied to the positive outcome of a model.
The following table shows a worked calculation for an automated resume screening tool.
| Group | Applicants | Selected | Selection rate | Impact ratio |
|---|---|---|---|---|
| Group A | 400 | 120 | 30.0% | 1.00 (reference) |
| Group B | 250 | 55 | 22.0% | 0.73 |
| Group C | 150 | 42 | 28.0% | 0.93 |
In this example, Group B has an impact ratio of 0.73, which is below the threshold, so the screening tool would be flagged for further investigation with respect to Group B. Group C passes the rule.
Legal origin and current status
The rule comes from the Uniform Guidelines on Employee Selection Procedures, adopted in 1978 by several United States federal agencies, including the Equal Employment Opportunity Commission. The guidelines describe it as a practical enforcement tool for deciding when a selection procedure merits attention under Title VII of the Civil Rights Act.
Courts treat the rule as persuasive but not binding. They often combine it with statistical significance testing and have accepted adverse impact claims with ratios above 0.8, as well as rejected claims with ratios below it when samples were small. The rule has gained a second life in algorithmic auditing: bias audit requirements for automated employment decision tools, such as New York City Local Law 144, require employers to publish impact ratios, although that law does not itself set a passing threshold.
The table below summarizes how the rule is used in different settings.
| Setting | Role of the rule | Binding force |
|---|---|---|
| Federal employment enforcement | Initial screen for adverse impact | Guideline, not statute |
| Litigation | One piece of statistical evidence | Weighed with other evidence |
| Algorithmic bias audits | Reported metric for each group | Reporting often required, threshold not |
| Internal model governance | Alert threshold in fairness dashboards | Set by company policy |
The rule compared with other fairness metrics
The four-fifths rule measures only outcome rates and ignores whether the selected people were actually qualified. Other fairness metrics in the AI ethics toolkit condition on the true outcome or on the model’s prediction, and they can lead to different conclusions about the same system.
| Criterion | What it compares | Uses ground truth labels |
|---|---|---|
| Four-fifths rule | Ratio of selection rates, threshold 0.8 | No |
| Demographic parity | Equality of selection rates | No |
| Equal opportunity | True positive rates across groups | Yes |
| Equalized odds | True and false positive rates | Yes |
| Predictive parity | Precision across groups | Yes |
A well-known result in fairness research shows that several criteria cannot hold at once when base rates differ between groups. A model can therefore pass the four-fifths rule while violating equalized odds, or the reverse, which is why audits usually report more than one metric.
Applying the rule to AI systems
Applying the rule to a machine learning model requires defining what counts as selection, which groups are compared and which data are used. For a ranking or scoring model, selection usually means passing a decision threshold or reaching the next stage of a process.
- Compute impact ratios at the actual operating threshold, because ratios change as the cutoff moves.
- Check every protected group and relevant intersections, such as sex combined with race, not only the largest groups.
- Use historical production data where possible, since test sets may not reflect the real applicant pool.
- Pair each ratio with a significance test such as a two-proportion z-test or Fisher’s exact test.
- Repeat the analysis after retraining, because model drift can change ratios over time.
When a ratio falls below 0.8, teams investigate causes such as proxy discrimination, representation bias in training data or features that correlate with protected attributes. Mitigation options include reweighting training data, adjusting thresholds or removing problematic features, each with its own legal and accuracy implications.
Limitations and common misreadings
The rule is sensitive to sample size. With small groups, one or two decisions can push a ratio across the threshold, so a single hire may decide whether a procedure appears compliant. Conversely, in very large samples, a ratio of 0.85 can reflect a statistically significant disparity that the rule does not flag.
Another limitation is that the rule says nothing about job-relatedness. A procedure with adverse impact can still be lawful if it is valid and necessary for the job, and a procedure that passes may still be unfair to individuals. The rule is also a group fairness measure and cannot detect unfair treatment of specific people within a group.
| Misreading | Why it is wrong | Better practice |
|---|---|---|
| A ratio above 0.8 proves fairness | The rule only screens outcomes | Combine with other metrics |
| A ratio below 0.8 proves discrimination | Small samples and validity matter | Test significance and job-relatedness |
| Removing protected attributes ensures compliance | Proxies still drive disparities | Measure outcomes directly |
| One audit is sufficient | Data and models change | Monitor ratios continuously |
Explained without jargon
The four-fifths rule is a quick fairness check. Suppose a hiring system picks 50 out of every 100 applicants from one group. Then it should pick at least 40 out of 100 from every other group. If it picks fewer, something may be wrong, and someone should look closer.
It works like a smoke alarm. An alarm that goes off does not prove there is a fire, and a silent alarm does not guarantee safety. It simply tells people when to investigate.
A typical use case
A mid-sized logistics company uses a machine learning model to rank applicants for warehouse supervisor roles, and candidates scoring above 70 are invited to interviews. Before renewing the vendor contract, the company commissions an independent bias audit covering 2,400 applications from the previous twelve months.
The auditors find that women are invited at a rate of 18 percent and men at 27 percent, an impact ratio of 0.67. A two-proportion test confirms that the gap is statistically significant. Further analysis shows that the model rewards years of forklift certification, a feature that tracks past hiring patterns more than supervisory ability.
The company removes the feature, retrains the model and validates the new version against supervisor performance ratings. The revised impact ratio rises to 0.91, predictive accuracy stays roughly unchanged, and the company schedules quarterly impact ratio monitoring as part of its model governance.