Bias & Fairness

About This Module

An AI system can be accurate overall and still treat some groups much worse than others. Algorithmic bias usually comes from a handful of sources: training data that reflects past discrimination, groups that are under-represented in the data, labels that measure the wrong thing, feedback loops that amplify early errors, and deliberate manipulation of the data. Bias on its own is not always harmful, but it becomes a problem when it leads to discrimination.

The lessons begin with these sources of bias, then look at a real case in facial analysis software. You then learn how fairness is measured, with demographic parity, equalized odds and related metrics, and how toolkits such as Fairlearn put them into practice. A well-known example is the COMPAS risk score, where ProPublica and the tool’s maker each pointed to a different, reasonable definition of fairness.

That example leads to the module’s central lesson: when groups have different base rates, several fairness definitions cannot all be satisfied at once, so choosing between them is a value judgment, not only a technical one. Bias can be reduced before training, during training or after it, but none of these steps replaces that judgment. Watch the lessons in order, then continue on to Transparency & Explainability.