Techniques applied during AI development to reduce or eliminate biases in models and datasets.
A suite of interventions - preprocessing (rebalancing or reweighting data), in-processing (fairness-aware learning objectives), and postprocessing (adjusting predictions to meet fairness criteria) - that systematically reduce unwanted disparities. Governance best practices include selecting mitigation strategies aligned to the organization’s risk tolerance and compliance needs.
A criminal-justice tool that predicts recidivism risk applies a bias-mitigation algorithm: during training, it adds a fairness penalty that reduces prediction gaps between white and Black defendants. After retraining, recidivism prediction rates are statistically equivalent across race groups, and the tool’s deployment guidelines are updated accordingly.
