Bias

An AI system can appear accurate overall while consistently producing worse results for certain people, situations, or types of data; this pattern is known as bias. It may enter a model through unrepresentative training examples, historical inequalities, labeling decisions, feature selection, optimization targets, or the environment in which the system is deployed. Bias can affect hiring tools, facial recognition, lending models, medical software, and other high-impact applications. Reducing it requires more than removing sensitive fields, because related variables may preserve the same patterns. Teams typically examine data coverage, compare performance across relevant groups, document limitations, test real-world outcomes, and involve domain experts and affected communities throughout development and monitoring.