Predictive Parity

Predictive parity (also called predictive value parity) requires that the positive predictive value of an AI hiring model — the probability that a positively classified candidate is truly qualified — be equal across all demographic groups. In practical terms, if an AI screening tool advances a candidate, the probability that the candidate will perform well on the job should be the same regardless of the candidate's demographic group. Predictive parity addresses a different concern than demographic parity. While demographic parity focuses on equal selection rates, predictive parity focuses on equal prediction accuracy. A model can satisfy demographic parity while violating predictive parity, and vice versa. The impossibility theorem (Chouldechova, 2017) proves that when base rates differ across groups, a model cannot simultaneously achieve both demographic parity and predictive parity. This tension is central to the fairness debate in AI hiring: employers must decide which fairness criteria to prioritize based on their values, legal requirements, and business context. In practice, regulatory compliance typically focuses on the four-fifths rule (a relaxed form of demographic parity), while predictive parity is examined as a best-practice quality check.