“We introduce the notion of equal opportunity as a nondiscrimination criterion for supervised learning.”
Moritz Hardt, Eric Price, and Nathan Srebro proposed equal opportunity in 2016 as a formal fairness criterion for supervised learning. For a beneficial outcome, the criterion asks whether qualified individuals have equal true-positive rates across protected groups. The paper also discussed equalized odds, which considers both true-positive and false-positive rates. These definitions helped make group disparities testable within a statistical framework. Its force comes from connecting a technical result to a wider question about how intelligent systems are built, evaluated, and placed in human institutions.
A fairness definition identifies which errors matter and for whom. In a lending example, equal opportunity might compare approval rates among people who would repay. In medical detection, it might compare how often illness is found among patients who truly have it. The choice of outcome and reference label is morally and empirically consequential. If historical labels are biased or incomplete, satisfying the equation may still reproduce injustice. This interpretation avoids treating intelligence as a single score. Performance depends on the task, data, environment, objective, and people who must live with the outcome.
Teams can use group metrics to uncover performance hidden by an overall average. They should report confidence intervals, examine smaller intersections, and investigate why disparities occur. Technical adjustments may change thresholds or predictions, but durable improvement can require better data, changed access, redesigned services, or a different policy. Fairness evaluation is strongest when paired with domain expertise and input from affected communities. The quotation offers a lens for current AI debates: ask what changed technically, what remained uncertain, and whether the surrounding organization can manage the consequences of error.
Multiple fairness criteria can be mathematically incompatible when base rates differ. There is no universal metric that resolves every ethical conflict. Organizations should disclose the definition they chose, explain why it fits the decision, and document who participated in that choice. The existence of a formula does not remove political responsibility; it makes some assumptions more visible and therefore more open to challenge. Good governance does not reduce ambition; it directs ambition toward results people can trust. Clear ownership, meaningful appeal, and honest reporting turn a clever system into accountable infrastructure. Hardt’s work gave fairness debates sharper tools. The next step is to use those tools transparently, while acknowledging the value judgments built into targets, labels, and tradeoffs. The most valuable legacy is a better question, not a larger slogan: what evidence would show that this idea improves human outcomes, and what evidence would tell us to stop? The work continues with care.
For a beneficial prediction, equal opportunity compares true-positive rates among members of different protected groups.
Equalized odds is broader because it considers both true-positive and false-positive rates across groups.
A fairness calculation depends on the outcome treated as ground truth, such as repayment, qualification, or diagnosis.
Historical labels may reflect unequal access or biased decisions, so satisfying a metric cannot validate the label automatically.
Different criteria can conflict, particularly when underlying outcome rates differ across the measured groups.
Institutions must justify their choice, involve affected people, disclose tradeoffs, and preserve meaningful routes for appeal.
Explore more "Quotes of The Day"
Continue exploring how writers, thinkers, artists, and advocates have described the choices that shape an everyday life.