Category: Fairness & Bias

Demographic Parity

Also known as: Statistical Parity, Acceptance Rate Parity, Benchmarking

Simply put

Demographic parity is a way of checking whether an AI model treats different groups similarly by comparing how often each group receives a positive outcome, such as a loan approval or a job offer. Under this measure, the model is considered fair if the rate of positive outcomes is roughly the same across groups defined by a sensitive characteristic. It focuses on whether the outcome depends on group membership, not on whether individual predictions are accurate.

Formal definition

Demographic parity is a group fairness metric requiring that a classifier's positive prediction rate be independent of membership in a sensitive group, so that the probability of a positive outcome is equal across groups (P(Y-hat = 1) equal for all groups). As commonly defined, it is also referred to as statistical parity, acceptance rate parity, and benchmarking, and it corresponds to the notion of no disparate impact in the sense of equal proportions of positive predictions per group. Note that this definition constrains only the distribution of predicted outcomes across groups and does not, by itself, account for differences in base rates or predictive accuracy; it should be distinguished from other fairness criteria such as equalized odds. The specific definitions and framing summarized here are drawn from the cited technical and educational sources and are not derived from any binding regulatory instrument.

Why it matters

Demographic parity gives practitioners a straightforward, outcome-focused lens for examining whether an AI model distributes positive decisions—such as loan approvals or job offers—at similar rates across groups defined by a sensitive characteristic. Because it constrains only the distribution of predicted outcomes and not the accuracy of individual predictions, it is often one of the first fairness checks teams reach for when assessing whether model outputs appear to depend on group membership. In many organizations, this makes it a useful diagnostic within broader efforts to identify potential disparate impact.

At the same time, its focus on equal positive-prediction rates is also its principal limitation, and this is where professionals frequently err. Demographic parity does not, by itself, account for legitimate differences in base rates between groups or for differences in predictive accuracy. Treating it as a sufficient or universal definition of fairness can obscure trade-offs with other criteria, such as equalized odds, which are not generally satisfiable simultaneously. It should therefore be understood as one metric among several, chosen deliberately for a given context rather than assumed to be the correct measure everywhere.

It is important to note that the definitions and framing summarized here are drawn from technical and educational sources, not from any binding regulatory instrument. While demographic parity is sometimes associated informally with the concept of "no disparate impact," that association is a conceptual parallel and should not be read as establishing that satisfying this metric meets any specific legal standard. Whether and how a fairness metric maps onto a jurisdiction's legal requirements is a separate, context-dependent question.

Who it's relevant to

Data scientists and model developers

Those building and evaluating classifiers use demographic parity as one group fairness metric among several, computing positive-prediction rates across sensitive groups. They should be aware that it does not account for base-rate or accuracy differences and generally cannot be satisfied at the same time as certain other criteria, making metric selection a deliberate design decision.

Model risk and validation teams

Professionals responsible for measuring and monitoring model risk may examine demographic parity as part of assessing whether outcomes appear to depend on group membership. They should treat it as a diagnostic rather than a complete fairness assessment, and distinguish it from validation of overall predictive performance.

AI governance and compliance specialists

Those establishing policies and oversight for AI systems may reference demographic parity when defining acceptable fairness criteria. They should note that the metric as defined here is not derived from any binding regulatory instrument, and that its informal association with "no disparate impact" does not establish compliance with any specific legal standard.

Auditors and legal professionals

Reviewers examining AI systems for potential disparate treatment of groups may encounter demographic parity in technical documentation. They should understand its scope and limitations—that it addresses only the distribution of positive predictions—and confirm separately how, if at all, it relates to the legal requirements applicable in the relevant jurisdiction.

Inside Demographic Parity

Group-level outcome comparison

Demographic parity, as commonly defined, is a group fairness criterion that compares the rate of a favorable model outcome (for example, a positive classification or approval) across groups defined by a protected or sensitive attribute. It is satisfied when that positive outcome rate is approximately equal across the groups being compared.

Independence between prediction and protected attribute

The criterion is typically framed as statistical independence between the model's predicted outcome and the protected attribute. In this framing, membership in a group should not change the probability of receiving the favorable outcome.

Protected or sensitive attribute definition

Applying the metric requires specifying which attribute (such as a legally protected characteristic in the relevant jurisdiction) defines the groups. The choice of attribute and how groups are constructed is a definitional input, not an output of the metric itself, and can materially affect results.

Tolerance or disparity threshold

Because exact equality of rates is rarely achieved in practice, assessments typically apply a tolerance band or ratio-based threshold to decide whether observed differences are treated as acceptable. The threshold is a chosen parameter and is not fixed by the metric definition.

Relationship to other fairness criteria

Demographic parity is one of several group fairness definitions and is generally distinct from criteria that condition on ground-truth labels (such as equalized odds or equal opportunity). These criteria can be mutually incompatible in many settings, so satisfying demographic parity does not imply satisfying others.

Common questions

Answers to the questions practitioners most commonly ask about Demographic Parity.

Does demographic parity mean a model is fair?

Not on its own. Demographic parity is one specific statistical fairness criterion—typically requiring that the rate of a favorable outcome be roughly equal across demographic groups—but it is not synonymous with fairness in general. Fairness is a broader, often normative and context-dependent concept, and multiple fairness criteria (such as equalized odds or calibration) can conflict with one another. Satisfying demographic parity does not guarantee that a system is fair across all relevant dimensions, and in some contexts it may not be the appropriate criterion at all.

Is demographic parity the same as removing bias from a model?

No. Bias and fairness are distinct concepts that experts avoid blurring, and demographic parity addresses only a particular parity-of-outcomes condition. A model can satisfy demographic parity while still exhibiting other forms of bias, and conversely a model with measurable bias in some sense may still be judged appropriate under a different criterion. Demographic parity is a measurable statistical property, not a comprehensive de-biasing guarantee.

How is demographic parity typically measured in practice?

It is commonly assessed by comparing the rate of a defined favorable outcome across groups defined by a protected or sensitive attribute, then examining whether those rates are equal or within a chosen tolerance. Practitioners generally must decide in advance which attribute defines the groups, what counts as the favorable outcome, and what deviation is treated as acceptable. Because these choices are analyst-defined rather than universally fixed, the same model can appear to satisfy or fail the criterion depending on how the measurement is set up.

What tolerance or threshold should be used when checking demographic parity?

There is no single universally mandated threshold. Some teams use an absolute difference in outcome rates, others use a ratio, and the acceptable range often depends on organizational policy, sector context, and any applicable legal or regulatory expectations. Given contested definitions and evolving regulatory treatment, the threshold and its justification should typically be documented and agreed with relevant oversight functions rather than assumed.

How does demographic parity relate to model validation and governance responsibilities?

Assessing demographic parity can form part of model validation activities and can also be a governance-level policy requirement, but the two roles are distinct. Validation typically involves independent testing of whether a chosen fairness metric behaves as intended, while governance concerns who sets the policy on which metric applies, who approves the tolerance, and how outcomes are escalated. Where these functions sit often maps to the lines-of-defense structure, and this entry does not prescribe a single arrangement.

What are common pitfalls when implementing demographic parity constraints?

Frequently cited pitfalls include treating it as the only fairness measure without considering criteria it may conflict with, enforcing parity of outcomes without examining whether that is appropriate for the use case, and applying it without clear documentation of the group definitions and thresholds used. Practitioners should also note that adjusting a model to meet demographic parity can affect other performance and fairness properties, so trade-offs generally need to be evaluated and recorded rather than assumed away.

Common misconceptions

Achieving demographic parity means the model is fair or non-discriminatory.

Demographic parity is a specific statistical criterion about equal positive outcome rates across groups. Fairness is a broader, context-dependent, and often contested concept. Satisfying one group fairness metric does not establish overall fairness, does not guarantee compliance with any particular jurisdiction's anti-discrimination law, and may conflict with other fairness definitions.

Demographic parity and equal accuracy or equal error rates across groups are the same thing.

Demographic parity concerns the rate of favorable outcomes and does not reference ground-truth labels, whereas criteria based on error rates (such as equalized odds or equal opportunity) condition on the true outcome. These are distinct measures and can point in different directions for the same model, so they should not be treated as interchangeable.

Demographic parity is a universally required or legally mandated standard.

As commonly defined, demographic parity is a technical fairness criterion, not a standard imposed uniformly across regulatory regimes. Whether, when, and how such a measure is expected varies by jurisdiction, sector, and use case, and its use should not be presented as a settled legal requirement.

Best practices

Document the specific fairness definition being used, and state explicitly that demographic parity measures equal favorable-outcome rates and does not, on its own, establish fairness or legal compliance.

Define and record the protected attribute, how groups are constructed, and the tolerance or ratio threshold used to judge disparities, treating each as a deliberate, reviewable choice rather than a default.

Assess demographic parity alongside other group fairness criteria (such as those conditioning on ground-truth outcomes) and note where these criteria conflict or cannot be simultaneously satisfied.

Involve compliance, legal, and domain stakeholders when selecting a fairness criterion, since the appropriate measure is context- and jurisdiction-dependent and may carry different implications across use cases.

Describe demographic parity results as one input into risk assessment that helps identify and manage potential disparate outcomes, not as a control that eliminates bias or discrimination risk.

Monitor group outcome rates over time as data and populations shift, and re-evaluate the chosen metric and thresholds if the operating context changes.