AI Glossary / Governance, ethics and safety / Advanced

Adverse impact ratio

Definition

An adverse impact ratio compares the selection rate for one group with the selection rate for a reference group. In the U.S. employment-guideline calculation described here, the reference is the group with the highest selection rate. The ratio helps identify differences in outcomes for further examination. It does not by itself explain the difference or establish whether a selection procedure is lawful.

What it means

Selection rate means the proportion selected within a group, not that group's share of everyone selected. Comparing rates accounts for differing group sizes. If one group has a 30% selection rate and the reference has a 50% rate, the ratio is 0.60: the first rate is 60% of the reference rate.
U.S. Uniform Guidelines use a four-fifths, or 80%, rule of thumb. A ratio below 0.80 generally indicates a disparity warranting attention under that guidance. It is not a universal fairness threshold or a legal definition. Statistical and practical significance, sample size and context also matter; a result above the threshold does not settle every concern.

How it works

Define the decision being examined, such as advancement to interview, and the population eligible for that decision. For each group, divide the number selected by the number considered. Then divide each rate by the chosen reference rate, documenting that reference explicitly.
Keep the same decision stage and observation period across the comparison. Missing demographic information, inconsistent candidate records and different treatment of withdrawals can change the denominator. A number calculated from incomparable populations can look precise while answering the wrong question.
For the highest-rate convention, defined ratios fall between zero and one. If nobody in any group is selected, the reference rate is zero and the ratio is undefined. Do not quietly replace that undefined result with a passing score.

Three practical examples

• A recruitment team compares interview-selection rates across groups for one role and period. It retains the underlying applicant and selection counts so reviewers can assess the ratio's stability.
• An internal promotion review compares advancement rates at a documented stage. The team checks how employees entered the candidate pool before interpreting differences at the final stage.
• An AI-tool evaluation compares a proposed screening system with the current process. It examines group selection rates alongside task validity and error patterns, rather than treating a single ratio as a complete assessment.

Think of it this way

Think of comparing the proportion of people passing through two gates. Counting how many pass is insufficient if one queue is much larger. You first calculate the fraction passing from each queue, then compare those fractions. The comparison shows a difference in outcomes; it does not tell you why the gates produced it.

Worked example

For one fictional selection stage, group A has 200 candidates and 100 selections, giving a selection rate of 100 / 200 = 0.50. Group B has 120 candidates and 42 selections, giving 42 / 120 = 0.35. Group C has 80 candidates and 32 selections, giving 32 / 80 = 0.40.
Group A has the highest rate and is the reference. B's adverse impact ratio is 0.35 / 0.50 = 0.70, or 70%. C's ratio is 0.40 / 0.50 = 0.80, or 80%. Under the stated less-than-80% screening convention, B falls below the threshold; C is exactly at it, rather than below it. Neither result alone resolves the full assessment.
For B, the absolute selection-rate difference is 50% − 35% = 15 percentage points. The relative shortfall is 1 − 0.70 = 30%. These are different descriptions of the same rates and should not be interchanged.
Dividing B's 42 selections by A's 100 selections would give 42%, which is not the impact ratio because their candidate pools differ. In this example, B would need 48 selections to have a 40% rate and a ratio of 0.80 with the other counts fixed. That arithmetic is not a recommendation to make individual decisions by demographic quota.

Illustrative case study

Fictional Cedar Services reviews an automated interview-screening process. Its dashboard reports a favorable combined ratio across several roles. A reviewer notices that applicant groups are distributed differently across those roles, which also have different selection rates.
The team examines role-specific results and the underlying counts. It also discovers that one report includes people who never reached the screening stage while another excludes them. It corrects the population definitions and preserves the earlier report so the change can be explained.
The revised analysis identifies a disparity for investigation. The team examines the procedure and relevant evidence with appropriate employment expertise rather than declaring either discrimination or compliance from the ratio alone. The case illustrates why population design and interpretation deserve as much attention as the division itself.

Common misconceptions and mistakes

Using selected counts instead of selection rates: unequal group sizes distort that comparison.
Reading 0.70 as a 70-percentage-point difference: it is a ratio of rates.
Treating 0.80 as a universal safe harbor: the cited guidance includes qualifications.
Ignoring small counts: a single selection can materially change a ratio.
Assuming a favorable overall result resolves every subgroup or stage: aggregation can conceal relevant differences.

Business use

Use the ratio as one documented part of a selection-process review. Store the counts, group definitions, decision stage, reference and time period. Investigate material disparities and examine statistical uncertainty and practical consequences. Legal interpretation depends on the applicable jurisdiction and facts; this U.S. employment example should not be treated as a worldwide compliance test or a substitute for qualified advice on an actual decision process.

People also ask

Is the reference always the largest group?

No; the convention here uses the highest selection rate.

Is the ratio the same as model accuracy?

No. It measures relative selection rates.

Does equal selection establish complete fairness?

No; other outcomes, errors and process concerns may remain.

Check your understanding

Why is B's ratio 70%, not 42%?

Answer: It compares rates after accounting for each group's candidate count.

What is B's absolute rate gap?

Answer: Fifteen percentage points.

What happens when every selection rate is zero?

Answer: The reference-based ratio is undefined.

Related terms

• Disparate impact
• Demographic parity
• Reference group
• Statistical significance
• Selection rate
• fairness assessment
• intersectional analysis
• selection procedure

Learn next

Create a small selection table with unequal group sizes. Calculate counts, rates, ratios and percentage-point differences, then explain which additional evidence you would need before drawing a substantive conclusion.

More in Governance, ethics and safety

• Adversarial loss - Adversarial loss is a training objective derived from a competition between models or model components.
• Adversarial search - Adversarial search is decision-making by exploring possible actions and the responses of other agents whose interests conflict with yours.
• Adversarial Training - Adversarial training is a model-training approach that uses deliberately challenging, permitted modifications of training inputs to improve resistance to...
• AI governance - AI governance is the system of decision rights, responsibilities, policies and oversight through which an organization directs its development, acquisition...
• AI principles - AI principles are high-level commitments intended to guide how artificial intelligence is developed, deployed and used.
• AI risk management - AI risk management is the ongoing work of identifying, assessing, responding to and monitoring risks associated with developing or using AI.

Sources

29 CFR 1607.4(D), current eCFR, impact information and four-fifths qualifications. https://www.ecfr.gov/current/title-29/subtitle-B/chapter-XIV/part-1607/section-1607.4 EEOC, Questions and Answers on the Uniform Guidelines, questions 10–12 on selection rates and impact ratios. https://www.eeoc.gov/es/node/130157 Fairlearn, Common fairness metrics, broader assessment context. https://fairlearn.org/main/user_guide/assessment/common_fairness_metrics.html The three-group calculations and Cedar Services case are original teaching illustrations.

Previous: Adversary argument Next: AED