Demographic parity is a machine learning fairness metric that requires a model to produce favorable outcomes at the same rate across demographic groups. Also called statistical parity or independence, it evaluates outcome rates rather than whether individual predictions are correct. It is useful for detecting unequal allocation patterns, but it should be assessed alongside data quality, model accuracy, and context-specific fairness goals.
Demographic Parity
What Is Demographic Parity in AI Fairness?
Demographic parity is a machine learning fairness metric that requires favorable outcomes at the same rate across demographic groups. Also called statistical parity or independence, it checks whether predictions are associated with sensitive attributes, such as age, gender, or race, rather than whether each prediction is correct.
A recommendation model meets demographic parity when groups receive recommendations equally often. It measures group patterns, not fairness for one person at a time.
How Demographic Parity Is Measured
The core question is whether each group receives a positive prediction equally often.
- Define the decision, such as approving an account or advancing a candidate.
- Identify groups and protected attributes that may lawfully be used for evaluation.
- Calculate each group’s positive prediction rate: positive predictions divided by all predictions for that group.
- Compare the rates or their ratio. The probability of a positive prediction should be the same for group A and group B.
- Check sample sizes and uncertainty, then document and monitor results as the model and data change.
Fairlearn’s fairness metric guidance provides a practical reference for these comparisons.
A Simple Demographic Parity Example
A job-candidate screening tool that recommends 40% of applicants in one gender group and 40% in another meets demographic parity for that attribute. This alone does not prove fairness: training data, “qualified” definitions, and accuracy may differ across groups.
Demographic Parity vs. Equalized Odds and Predictive Parity
These metrics ask different questions. The right choice depends on the decision, outcome-label reliability, and potential harm.
| Metric | What it equalizes | Uses true outcomes? | Useful when |
|---|---|---|---|
| Demographic parity | Positive prediction rates | No | Checking unequal access or allocation |
| Equalized odds | Error rates for each true outcome | Yes | Both missed positives and false positives matter |
| Predictive parity | Accuracy among positive predictions | Yes | A positive prediction must have comparable meaning across groups |
Benefits of Demographic Parity
- It reveals unequal allocation patterns in AI-assisted decisions.
- It provides a baseline for monitoring group outcomes over time.
- It can trigger deeper reviews of data, rules, and model behavior before release.
Practical Limits of Demographic Parity
Equal selection rates are evidence to investigate, not a complete fairness verdict.
- It does not show whether predictions are accurate or justified for each group.
- Different base rates and flawed historical labels can make equal rates hard to interpret.
- Small groups can produce unstable rates, so teams should report sample sizes and uncertainty.
- Protected-attribute data requires lawful collection, strict access controls, and secure retention.
Using Demographic Parity When Building AI Applications
Build fairness checks into requirements, representative test datasets, release reviews, monitoring dashboards, audit logs, and human escalation paths. Teams using an AI app builder can specify checks alongside decision rules and test outputs by group before deployment. Emergent helps teams create and iterate on AI-powered applications, while builders remain accountable for choosing suitable metrics, protecting data, and reviewing real-world outcomes.
Start Building
on Emergent today
on Emergent today
Start Building
