In Depth
Algorithmic fairness addresses the problem of AI systems producing discriminatory outcomes, even when they are not explicitly designed to consider protected attributes. Bias can enter through training data (historical discrimination encoded in data), feature selection (using proxies for protected characteristics), or model design (optimizing for aggregate accuracy that masks disparities across groups).
Multiple mathematical definitions of fairness exist, and they are often mutually incompatible. Demographic parity requires equal prediction rates across groups. Equal opportunity requires equal true positive rates. Calibration requires equal accuracy within predicted risk levels. Choosing which fairness criterion to optimize involves value judgments about what constitutes equitable treatment in a specific context.
Practical fairness work involves measuring disparities across groups (disaggregated evaluation), investigating root causes, applying mitigation techniques (pre-processing data, in-processing constraints, post-processing adjustments), and monitoring deployed systems for drift. Tools like AIF360, Fairlearn, and Aequitas help practitioners measure and mitigate bias. However, technical solutions alone are insufficient; organizational commitment to equity and diverse teams that can identify potential harms are equally important.