Equal opportunity is a group fairness criterion for classification models which requires that people who truly deserve the positive outcome have the same chance of receiving it regardless of their protected group. Formally, it demands equal true positive rates across groups defined by a protected attribute such as sex, race or age. In AI ethics, it is one of the most widely cited ways to turn the general idea of non-discrimination into a measurable property of a model.
The criterion was formalized in 2016 by Moritz Hardt, Eric Price and Nathan Srebro as a relaxation of equalized odds. It focuses on the qualified population only, which makes it attractive in settings such as lending, hiring and medical screening, where the main harm is wrongly denying a benefit to someone who should have received it.
How equal opportunity is defined
Equal opportunity compares model behavior only among positive cases, meaning people whose true label is the favorable one: a borrower who repays, a candidate who succeeds, a patient who has the disease. Within that group, the share that the model correctly predicts as positive is the true positive rate, also called recall or sensitivity.
Written as a condition, the predicted label must be independent of the protected attribute given a positive true label: P(Ŷ = 1 | Y = 1, A = a) is the same for every group a. Because exact equality rarely occurs in finite data, practitioners report the true positive rate gap, the largest difference between groups, and set a tolerance such as a few percentage points.
The table below lists the quantities that enter the calculation and how each one is obtained.
| Quantity | Meaning | How it is computed |
|---|---|---|
| True label Y | Actual outcome, such as repayment | Observed later or from records |
| Predicted label Ŷ | Model decision after the threshold | Score compared with a cutoff |
| Protected attribute A | Group membership being assessed | Collected or inferred for auditing |
| True positive rate | Share of positives predicted positive | TP divided by (TP + FN) per group |
| TPR gap | Size of the fairness violation | Maximum minus minimum group TPR |
Equal opportunity compared with other fairness metrics
Equal opportunity is one member of a family of group fairness metrics, and each member encodes a different moral claim. The key difference is which errors the metric cares about: equal opportunity constrains false negatives among qualified people but places no condition on false positives.
| Criterion | What must be equal across groups | Typical concern addressed |
|---|---|---|
| Equal opportunity | True positive rate | Qualified people wrongly denied |
| Equalized odds | True and false positive rates | Both kinds of classification error |
| Demographic parity | Rate of positive predictions | Unequal selection rates overall |
| Predictive parity | Precision of positive predictions | Equal meaning of a positive score |
| Disparate impact ratio | Selection rate ratio, often 0.8 or above | Legal screening under the four-fifths rule |
Unlike demographic parity, equal opportunity respects differences in base rates: if one group truly has more qualified members, the model may select more of them. This makes it easier to justify where the true labels are trustworthy, and harder to justify where those labels themselves reflect historical bias.
Achieving equal opportunity in a model
Equal opportunity can be pursued at three stages of the machine learning pipeline, and each stage has different costs in accuracy, complexity and legal acceptability. The original proposal used post-processing with group-specific thresholds, which is simple but requires using the protected attribute at decision time.
| Stage | Example method | Main trade-off |
|---|---|---|
| Pre-processing | Reweighting or resampling training data | Indirect effect on the TPR gap |
| In-processing | Training with a TPR gap penalty or constraint | More complex optimization |
| Post-processing | Separate decision thresholds per group | Uses the attribute at decision time |
Because enforcing the constraint usually changes which cases are approved, there is often a fairness-accuracy trade-off. The size of that trade-off depends on the data: in some problems the TPR gap can be closed at almost no cost, while in others it noticeably reduces overall accuracy or raises the false positive rate for one group.
Where equal opportunity fits and where it does not
Equal opportunity suits decisions in which the positive outcome is a benefit and missing a deserving person is the most serious harm. It fits poorly where false positives impose heavy burdens, as in fraud flags or risk scores used against people, because the criterion is silent about who is wrongly accused.
| Situation | Is equal opportunity suitable | Reason |
|---|---|---|
| Loan approval with reliable repayment data | Usually yes | Protects creditworthy applicants |
| Disease screening referral | Usually yes | Missed cases are the main harm |
| Fraud or risk flagging | Usually not alone | Ignores false accusations |
| Hiring with biased past ratings | Only with caution | Labels may encode discrimination |
| No access to true outcomes | Not measurable | TPR requires ground truth |
Limitations and common misconceptions
The most important limitation is dependence on the true label. If past decisions shaped who appears qualified, for example when only approved applicants ever had the chance to repay, then equal true positive rates may simply reproduce measurement bias under a fair-looking number. Selective labels of this kind are common in lending and criminal justice.
- It is not individual fairness: two similar people can still receive different outcomes while group rates match.
- It cannot hold together with every other metric: when base rates differ, equal opportunity and predictive parity generally conflict for an imperfect classifier.
- Removing the protected attribute does not achieve it: fairness through unawareness leaves proxies in the data.
- Small groups give unstable estimates: a TPR gap from a few dozen positives needs confidence intervals before conclusions are drawn.
Equal opportunity is therefore best read as one lens in an AI audit, reported alongside false positive rates, calibration and qualitative review of how labels were produced, rather than as proof that a system is fair.
Explained without jargon
Imagine a bank using software to decide who gets a loan. Equal opportunity asks a simple question: among people who would actually repay, does the software approve them at the same rate in every group? If it approves 85 percent of reliable men but only 70 percent of reliable women, it fails the test.
The idea does not demand that every group receives the same number of loans. It only demands that deserving people are not overlooked more often because of who they are. The weak point is that it trusts past records about who was deserving, and those records can themselves be unfair.
A typical use case
A regional hospital deploys a model that flags patients for a follow-up diabetes screening. Before launch, its data science team audits the model on two years of records in which the actual diagnosis is known. Among patients who truly developed diabetes, the model flags 82 percent of one ethnic group but only 69 percent of another, a TPR gap of 13 percentage points.
The team traces part of the gap to fewer recorded lab results for the second group and adds features that do not depend on visit frequency. It then applies group-specific thresholds to close the remaining gap, accepting a small rise in unnecessary referrals. The revalidated model reaches a gap under 3 percentage points, and the hospital ethics board approves deployment with a requirement to recheck the metric every quarter.