How a model performs differently across demographic or contextual groups, used to measure fairness and bias.