Demographic parity

Demographic parity is a formal criterion of algorithmic fairness that requires equal rates of positive classification across protected demographic groups. If a decision system approves 30% of applicants from one group, it must approve approximately 30% from all other groups, regardless of the underlying distribution of qualifications or risk.

The criterion is motivated by a straightforward intuition: if the algorithm produces different outcomes for different groups, something has gone wrong. But the simplicity is deceptive. Demographic parity can force discriminatory treatment of individuals when group base rates differ. If Group A has more qualified applicants than Group B, demographic parity requires that some qualified applicants from A be rejected to maintain parity with B, or that less qualified applicants from B be approved.

The tension between group-level parity and individual-level merit is one of the central trade-offs in fairness research. Demographic parity treats the group as the unit of fairness; other criteria like equalized odds treat the individual's error experience as the unit. The choice between them is not mathematically decidable — it is a normative commitment that the mathematics formalizes but does not resolve.\n