What is fairness in AI?
Fairness in AI is the property that a model’s outputs do not systematically disadvantage people based on protected characteristics such as race, sex, age, disability, or national origin. A model can be unfair without ever seeing a protected attribute, because other inputs such as zip code, employment gaps, or purchase history stand in for it. Fairness work therefore starts by measuring outcomes across groups, whatever the model was fed.
Fairness is a family of definitions rather than a single test, and the definitions cannot all be satisfied at once when the groups differ in their underlying rates. Choosing which definition applies is a business and legal decision, made before the model is built.
Types of fairness
- Demographic parity. Each group receives the favorable outcome at the same rate. A loan model approves the same share of applicants from every group.
- Equal opportunity. Among people who truly qualify, each group is approved at the same rate. This definition focuses on false negatives.
- Equalized odds. Both the false positive rate and the false negative rate match across groups.
- Calibration across groups. A score of 0.8 means the same probability of the outcome for everyone, whichever group they belong to.
- Individual fairness. Two people who are alike in every relevant way receive the same result.
Fairness in practice
A regional bank builds a model to prioritize small business loan applications for faster review. The compliance team selects equal opportunity as the standard, since the harm they most want to avoid is a qualified owner waiting weeks. Testing shows applicants from rural counties are prioritized less often despite similar repayment history, traced to a feature based on distance from a branch. Removing the feature costs a little accuracy and closes the gap. The analysis and the decision are documented for the bank’s fair lending review.
What to watch for
- Measuring fairness usually requires the protected attributes you are trying not to use, held separately for testing under strict access control.
- A model that is fair on historical data can become unfair as the population or the process changes, so test after deployment.
- Vendors who claim a model is “bias-free” should be asked which definition they used and on whose data.
- Fairness analysis belongs in bias detection and fairness auditing as a repeatable process, with results kept for regulators.