When different groups of people (or models) answer the same test item differently despite having equal ability.