Is a fairness constraint the same as a fairness metric?

No. The metric measures something; the constraint specifies a condition that measurement should satisfy.

Does a soft fairness penalty guarantee a bound?

Not generally. Its tradeoff with the prediction objective can leave violations.

Can a model meet one constraint and still be unfair?

Yes. A constraint captures a limited concern and can miss other harms or depend on flawed data.

AI Glossary / Governance, ethics and safety / Intermediate

Fairness constraint

Definition

A fairness constraint is a formally specified condition that an AI model or decision process is required to satisfy with respect to a chosen fairness criterion. It may limit differences in selection rates, error rates, or another defined quantity across groups. The constraint translates one concern into a measurable requirement; satisfying it does not establish fairness in every ethical, social, or legal sense.

What it means

An ordinary prediction objective might seek the lowest average error. Adding a fairness constraint restricts which solutions are considered acceptable under a particular definition. For example, a team might limit a specified difference in true positive rates across named groups. The choice of groups, metric, conditioning variables, tolerance, and population is part of the requirement. These details cannot be left implicit. Different fairness criteria express different concerns, and a suitable criterion in one setting may be unsuitable in another. The metric should follow an analysis of the decision and its harms, not simply the default offered by a software library.

How it works

A constrained optimization formulation minimizes a loss while requiring fairness-related quantities to remain within specified bounds. Some methods change training through weighted classification problems; others adjust decisions after a model is trained. A soft penalty instead adds a cost for violating a fairness objective and may allow violations when another improvement outweighs that cost. These formulations are related but not interchangeable. Estimated constraints depend on the available sample, and satisfying them on training data does not guarantee satisfaction in deployment. A method may also find no useful feasible solution under a chosen model family and set of requirements.

Three practical examples

A research prototype limits differences in false negative rates across specified groups while evaluating the resulting prediction errors.
A service-allocation model is assessed against a defined bound on selection-rate disparity, alongside checks of eligibility labels and service quality.
A model-development team compares candidate solutions under several tolerances to understand how its chosen requirement affects performance. These are illustrative technical formulations, not recommendations that any one metric should govern all decisions.

Think of it this way

Think of a fairness constraint as one design requirement on a bridge, such as a maximum permitted deflection. Meeting that requirement matters, but it does not prove the bridge satisfies every engineering need. The analogy is limited because fairness criteria involve contested values and social context, not merely physical measurement. Deciding which disparity matters and what tolerance is acceptable requires judgment and participation beyond solving an optimization problem.

Worked example

Suppose a hypothetical project defines its constraint as an absolute difference of no more than 5 percentage points between the true positive rates of Groups A and B. There are 100 actual positive cases in each group. Candidate X correctly identifies 90 in A and 80 in B, giving rates of 90% and 80%, a 10-point gap that violates the requirement. Candidate Y identifies 86 in A and 83 in B, giving a 3-point gap that meets it on this sample. This calculation does not show which candidate has better overall accuracy because false positives are not provided. It also does not establish population-level compliance or justify the labels. The bound here compares the two groups directly; some software instead bounds each group's deviation from an overall mean, so the same numerical setting need not mean the same constraint.

Illustrative case study

Fictional case: Linden Scheduling studies an AI system for prioritizing service requests. The team initially selects a parity constraint because it is easy to compute. Discussions with service users reveal that missed urgent requests, rather than selection rates alone, are a central concern. Analysts revise the assessment, inspect label quality, and document remaining harms that the revised constraint cannot capture. This invented case illustrates that selecting a fairness constraint is a governance decision supported by technical analysis, not a purely mathematical choice.

Common misconceptions and mistakes

Removing sensitive attributes does not guarantee that a fairness constraint will be satisfied; other features can carry related information. Equal rates do not automatically mean equal treatment quality or equal access to the system. A constraint evaluated on broad groups may miss differences within intersecting groups. A numerical tolerance is not automatically a legal threshold or a scientific constant. Constraints can also conflict with one another or with other objectives under particular conditions. Such conflicts should be analyzed explicitly rather than hidden behind an overall fairness score.

Business use

Document the reason for each constraint, its exact formula, the affected groups, the data needed, and who approves changes. Use appropriate access and privacy controls for sensitive group information. Evaluate both constraint estimates and prediction performance on independent data, with uncertainty and sample sizes. Inspect operational harms that the formula leaves out, including exclusions before the model receives a case. Monitor after deployment because population and workflow changes can alter the measured disparity. If the requirement cannot be met with an acceptable service, reconsider the system or process rather than silently relaxing the bound.

Check your understanding

A model satisfies a five-point training-data disparity bound. May the team describe it as universally fair?

Answer: No. It has satisfied a specified empirical condition. Independent evaluation, uncertainty, deployment context, and concerns outside that condition still need attention.

Related terms

Learn next

Write one proposed fairness constraint in plain language and as a precise inequality. State its denominator, population, and tolerance. Then identify two plausible harms it would not detect and design additional evaluation for those gaps.

Sources

Agarwal and colleagues, A Reductions Approach to Fair Classification: https://proceedings.mlr.press/v80/agarwal18a.html Fairlearn, reductions and exact constraint definitions: https://fairlearn.org/main/user_guide/mitigation/reductions.html Fairlearn, fairness assessment guidance: https://fairlearn.org/main/user_guide/assessment/perform_fairness_assessment.html

Previous: FAIR principles Next: Fairness metric