Demographic parity, also called statistical parity, is a group fairness criterion that requires a model to assign the positive outcome at the same rate in every protected group, regardless of whether individuals in those groups differ in the true outcome. In AI ethics, it is the simplest and most widely cited way to formalize the idea that a decision system should not favor one demographic group over another.
Formally, a classifier satisfies demographic parity when the probability of a positive prediction is independent of the protected attribute. The criterion looks only at decisions, never at ground-truth labels, which is both its main strength and its most debated weakness.
How demographic parity is defined and measured
Demographic parity holds when P(Ŷ = 1 | A = a) is equal for every group a, where Ŷ is the model’s decision and A is the protected attribute, such as sex, ethnicity or age band. The quantity P(Ŷ = 1 | A = a) is often called the selection rate or positive rate of group a.
Exact equality almost never occurs in practice, so practitioners measure how far the system is from parity and apply a tolerance. The table below lists the usual measures.
| Measure | Formula | Parity value |
|---|---|---|
| Statistical parity difference | Selection rate of group A minus group B | 0 |
| Disparate impact ratio | Lower selection rate divided by higher | 1 |
| Maximum group gap | Highest minus lowest rate across all groups | 0 |
| Four-fifths threshold | Disparate impact ratio of at least 0.8 | Pass or fail |
The four-fifths rule comes from employment selection guidance in the United States and treats a ratio below 0.8 as evidence of adverse impact. It is a screening heuristic, not a statistical test, so small samples can produce ratios below 0.8 by chance and ratios above it can still hide meaningful disparities.
How demographic parity compares with other fairness metrics
Demographic parity differs from other fairness metrics because it ignores the true labels entirely. Criteria such as equalized odds, equal opportunity and predictive parity condition on the actual outcome or on the prediction, so they treat differences in base rates as legitimate.
| Criterion | What must be equal across groups | Uses true labels |
|---|---|---|
| Demographic parity | Rate of positive decisions | No |
| Equal opportunity | True positive rate | Yes |
| Equalized odds | True positive and false positive rates | Yes |
| Predictive parity | Precision of positive decisions | Yes |
These criteria generally cannot all hold at once. Well-known impossibility results show that when base rates differ between groups, a classifier that is not perfect cannot satisfy demographic parity together with equalized odds or calibration-based criteria. Choosing a metric is therefore a normative decision, not a technical default.
Demographic parity is also distinct from individual fairness, which asks that similar individuals receive similar treatment. A system can meet parity at the group level while treating comparable individuals inconsistently within each group.
Achieving demographic parity in a model
Demographic parity can be enforced at three points in the machine learning pipeline, and each point carries different costs and legal implications. The table summarizes the standard bias mitigation families.
| Stage | Typical technique | Main trade-off |
|---|---|---|
| Pre-processing | Reweighting or resampling training data | Changes data the model learns from |
| In-processing | Fairness constraint or penalty in training | Usually reduces measured accuracy |
| Post-processing | Group-specific decision thresholds | Uses the protected attribute at decision time |
Simply removing the protected attribute, an approach called fairness through unawareness, rarely achieves parity. Other features such as postal code, school or purchase history act as proxies for the protected attribute, so the model reconstructs group membership and the gap persists.
Post-processing with group-specific thresholds is the most direct route to parity, but in employment, lending and housing it can conflict with anti-discrimination law that prohibits explicit use of protected characteristics in individual decisions. Legal review should precede any such deployment.
When demographic parity is the right criterion
Demographic parity is most defensible when the historical labels themselves are suspected of bias, so matching them is not a meaningful goal. If past hiring decisions reflected discrimination, a model that reproduces those labels accurately would also reproduce the discrimination.
| Situation | Is parity appropriate | Reason |
|---|---|---|
| Labels reflect past discrimination | Often yes | Accuracy would encode historical bias |
| Allocating scarce opportunities | Often yes | Equal access is the stated policy goal |
| Ad delivery and recommendations | Sometimes | Exposure gaps cause allocative harm |
| Medical diagnosis | Usually no | True prevalence differs between groups |
| Fraud or credit default prediction | Debated | Base rates and law vary by context |
In domains where true prevalence genuinely differs, forcing equal positive rates means either missing real cases in one group or flagging healthy individuals in another. Here, error-based criteria such as equal opportunity usually reflect the ethical goal more closely.
Limitations and common misconceptions
Demographic parity is a necessary-looking but insufficient guarantee, and several misunderstandings recur in audits and policy discussions.
- Parity is not accuracy: a model can reach equal selection rates by choosing randomly within one group, which satisfies the metric while serving that group poorly.
- Parity can be gamed by thresholds: adjusting cut-offs changes rates without addressing the underlying measurement bias in features.
- Aggregate parity can hide subgroup gaps: rates may be equal for women and men overall but unequal for intersectional subgroups.
- Small groups produce noisy estimates: confidence intervals should accompany every reported rate.
- Parity at one stage does not ensure parity downstream: a fair screening model can feed a biased human or automated decision later.
Because of these limits, demographic parity works best as one indicator within an AI audit, reported alongside error-based metrics, subgroup analysis and a documented rationale for the chosen fairness definition.
Put simply
Demographic parity asks a simple question: does each group get a yes from the system at the same rate? If a loan model approves 40 percent of one group, it should approve about 40 percent of every other group.
This sounds fair, and often it is. But it does not check whether the yes decisions were correct. Equal rates are a useful warning light, not proof of fairness.
In practice
A regional employer uses a screening model to decide which of 2,000 applicants for warehouse supervisor roles move to an interview. An internal fairness review finds that the model advances 30 percent of male applicants but only 21 percent of female applicants, a disparate impact ratio of 0.70, below the four-fifths threshold.
Analysis shows that the model relies heavily on years of forklift certification, a feature that reflects past assignment practices rather than supervisory ability. The team removes it, reweights the training data and validates the new model on a held-out set.
The revised model advances 28 percent of men and 25 percent of women, a ratio of about 0.89. The team also confirms that interview-to-offer rates are similar for both groups, so the improvement in parity does not simply move the gap to a later stage.