A fairness criterion requiring that a model’s positive prediction rate is equal across demographic groups defined by protected attributes such as age, gender, or ethnicity. For agencies deploying AI in targeting and personalization, demographic parity is one measure of whether the system is treating different audiences equitably.
Also known as statistical parity, group parity, demographic fairness
Demographic parity holds that a model satisfies fairness requirements when the proportion of positive predictions is the same across demographic groups. If a loan application model approves 70% of applications from group A but only 40% from group B, it violates demographic parity even if the model’s overall accuracy is high. The disparity may reflect historical discrimination embedded in the training data, or it may reflect genuine differences in the underlying distribution of the relevant feature, which complicates the determination of whether a disparity is a fairness problem.
Demographic parity is one of several formal fairness definitions, and they are mathematically incompatible with each other in most real-world settings. Satisfying demographic parity may require violating equalized odds (equal true positive and false positive rates across groups) or predictive parity (equal precision across groups). Choosing which fairness criterion to optimize for is a values decision, not a technical one.
Regulatory frameworks including the EU AI Act and sector-specific fair lending laws treat disparate impact, which is closely related to demographic parity violations, as a compliance concern. Responsible AI guidelines increasingly require fairness audits that measure parity across relevant demographic dimensions before high-stakes systems are deployed.
Agencies use and recommend AI systems that make predictions about individuals: who sees which ads, which leads get scored as high-value, which customers receive which personalization treatments. Each of these is a decision that could produce disparate outcomes across demographic groups, and each could expose the agency or its clients to regulatory scrutiny or reputational risk.
Ad targeting systems have built-in disparity risks. Algorithmic ad delivery does not simply show ads to whoever the advertiser targets; it optimizes delivery to the users most likely to engage, which can produce demographic concentration effects even when the advertiser did not intend to target a specific group. Agencies responsible for fair housing, employment, or credit advertising need to actively audit delivery demographics, not just targeting inputs.
Training data encodes historical disparities. A model trained on historical conversion data will reproduce the conversion patterns in that data, including any disparities that resulted from past practices that favored certain demographic groups. Bias in AI is often the mechanism by which historical inequities get systematized at scale in a way that is harder to see and easier to dismiss as “what the data says.”
Fairness auditing is increasingly a client deliverable. Enterprise clients in financial services, healthcare, and consumer sectors are beginning to require fairness documentation as part of AI vendor due diligence. Agencies that can conduct and present demographic parity analyses are better positioned for those conversations than agencies that treat fairness as someone else’s problem.
An agency builds a content personalization model for a consumer financial services client. Before deployment, a fairness audit measures the positive recommendation rate, which is defined as the model recommending a premium product, across age groups. The audit finds that the recommendation rate for users over 60 is 18%, versus 34% for users aged 25-44. Investigation reveals that the training data underrepresents older users who converted during the period when the product was heavily promoted to younger audiences. The agency rebalances the training data by demographic cohort and reruns the audit before deployment. The client’s legal team signs off on the revised parity metrics as part of the launch approval process.
The governance and disclosure module of the workshop covers the internal standards your agency needs to use AI without creating fairness risks for your clients or their customers.