In AI and cyber security, discrimination is unequal treatment that may occur when a system produces different outcomes for different groups. It can result from biased training data, design choices, proxy variables, or how people use the system. A model can have strong overall accuracy yet perform less reliably for a particular group.
Discrimination matters when automated systems influence decisions about access, priority, or services, including healthcare triage. Before deployment, teams can compare performance and outcomes across relevant groups, investigate disparities, and document mitigations and remaining risks. These checks help identify potential harm; a disparity alone does not establish its cause or prove unlawful discrimination.