Is this identical to equalized odds?

No. Equalized odds also requires matching false-positive rates in the binary setting.

Does the positive label always mean a benefit?

No. A positive label can denote a condition such as a defect. The interpretation must be explicit, especially when the criterion is discussed as opportunity.

Can equality of opportunity certify fairness?

No. It addresses one statistical relationship and depends on labels, groups, evidence, and the decision context.

AI Glossary / Governance, ethics and safety / Intermediate

Equality of opportunity

Definition

In binary machine-learning fairness, equality of opportunity usually means that the true-positive rate is the same across the groups being compared. Among cases that genuinely have the positive target label, each group has the same rate of receiving a positive prediction. This technical criterion is narrower than the broader ethical and political idea of equal opportunity, and depends on how the target label is defined.

What it means

Imagine a system intended to identify requests that qualify for additional assistance. The criterion asks whether qualifying requests are recognized at similar rates across groups. It does not require the same overall selection rate, because groups may contain different proportions of positive-labeled cases. It also leaves false-positive rates unconstrained. The criterion therefore answers a specific error-distribution question. Satisfying it does not prove that the labels are fair, that the opportunity itself is accessible, or that the surrounding decision process treats people appropriately.

How it works

For each group, count the true positives: positive-labeled cases receiving a positive prediction. Divide by all positive-labeled cases in that group, including those the model missed. The result is the true-positive rate, also called recall or sensitivity in relevant contexts. Compare these rates, with attention to uncertainty and sample size. Exact population equality is the mathematical ideal; an empirical assessment must state its tolerance and evidence. Model changes or threshold adjustments may alter the gap, but their effects on other errors and the actual decision need evaluation. Group-dependent interventions also require context-specific review rather than automatic adoption.

Three practical examples

A service team examines whether requests independently confirmed to need specialist support are identified equally often across communication channels or relevant user groups.
A training platform checks whether learners meeting a clearly validated readiness criterion receive a positive recommendation at similar rates across groups.
A quality-control system compares how often genuinely defective items from different production settings are flagged. The statistical form is the same, though the meaning and consequences of the positive label differ.

Think of it this way

Think of two search teams looking for known eligible cases in separate collections. Equality of opportunity compares what fraction of the eligible cases each team finds. It does not compare everything the teams select, because they might also select ineligible cases at different rates. The analogy is limited because real labels can be uncertain or biased, and identifying an eligible case is not the same as delivering a meaningful opportunity.

Worked example

Suppose Group A contains one hundred positive-labeled cases and Group B contains fifty. A classifier correctly identifies eighty in A and forty in B. The true-positive rates are 80/100 and 40/50, both 80 percent. On this sample, the equality-of-opportunity rates match, even though the counts differ. Now suppose each group also contains one hundred negative-labeled cases. The classifier incorrectly selects ten in A and thirty in B. False-positive rates are 10 and 30 percent, so equalized odds does not hold. Overall selection rates are 90/200 = 45 percent for A and 70/150, about 46.7 percent, for B. These different quantities are not interchangeable. The example demonstrates which condition is satisfied; it does not establish that the system is fair overall or that the small sample proves population equality.

Illustrative case study

Fictional case: Rowan Learning reviews a recommendation tool intended to identify people ready for an advanced workshop. The team first checks whether its readiness label reflects the workshop's actual prerequisites. It then compares true-positive rates and discovers that one group's eligible participants are missed more often. Proposed changes are assessed alongside incorrect recommendations, accessibility, and available places. The fictional case shows how a metric can focus an investigation while leaving broader questions open; it does not prescribe a selection policy or claim a real improvement.

Common misconceptions and mistakes

Equal true-positive rates do not mean equal numbers selected or equal overall selection rates. They also do not imply equal precision, since false positives and base rates can differ. A poorly chosen positive label can make the calculation look reassuring while preserving a harmful definition of qualification. Another mistake is treating a tiny observed gap as meaningful without uncertainty analysis, or treating a zero gap in a small sample as proof of exact equality. The mathematical criterion cannot replace review of the opportunities, resources, and institutional conditions surrounding the model.

Business use

State what a positive outcome means and why its label is a credible basis for assessment. Examine the positive-labeled sample available for each relevant group and disclose missing or selectively observed outcomes. Report true-positive rates with counts and uncertainty, plus other error measures needed for the decision. Discuss the criterion with domain specialists and affected people before treating it as an objective. Evaluate proposed mitigations for utility, other harms, and practical constraints. Continue monitoring after deployment because changing populations and workflows can alter the observed rates.

Check your understanding

A model finds forty of fifty positive cases in one group and eighty of one hundred in another. Are their true-positive rates equal?

Answer: Yes. Both are 80 percent. The different counts do not create a rate difference, but other errors and uncertainty still need review.

Related terms

Learn next

Build a confusion matrix for each relevant group in a small evaluation set. Calculate true-positive rates using only actual positive cases as denominators. Then calculate false-positive rates to see what the equality-of-opportunity criterion leaves unspecified.

Sources

Hardt, Price, and Srebro, Equality of Opportunity in Supervised Learning, original paper, definition 2.2: https://www.cs.utexas.edu/~ecprice/papers/fairness.pdf Fairlearn, Common fairness metrics: https://fairlearn.org/main/user_guide/assessment/common_fairness_metrics.html

Previous: Equal impact Next: Equality symbol