Fairness (AI)

There is no single test for fairness, because people disagree about what it requires. Common measures include equal treatment of similar individuals, similar approval rates across groups, and similar error rates across groups. These can conflict, so a team must choose the measure that fits the setting and record why. Which one fits can depend on the setting, for example lending or hiring. Fairness work runs through the AI lifecycle: checking the data, testing outputs by group, adjusting thresholds and monitoring in use. It also depends on context outside the model, such as whether a decision can be appealed. Beyond AI, fairness is a wider principle in conduct and consumer law, as in treating customers fairly. In practice a fairness report names the comparison chosen and the gap found, so a committee can accept the result or send the model back.

An insurer tests a pricing model and finds that customers in one postcode area pay noticeably more, though their claims history is no worse. The postcode is acting as a stand-in for ethnicity. The insurer compares prices and claim outcomes across groups, drops the variable, and adds a quarterly fairness report to its model committee.