demographic parity
/ dem-uh-GRAF-ik PAIR-uh-tee /
Demographic parity is one specific way to define a "fair" algorithm: it says the rate of positive decisions should be the same across groups. If a hiring filter advances 30% of male applicants, demographic parity demands it also advance 30% of female applicants — and 30% of every other group. It cares only about the bottom line of how many got the good outcome, not about who, individually, deserved it.
The appeal is intuitive: if a screening tool lets 5% of one group through but 25% of another, something looks off, even before you investigate why. The catch is that demographic parity is blind to legitimate differences. Suppose, for a fair reason, two groups genuinely differ in some relevant trait — say, average years of relevant experience in an applicant pool. Forcing identical acceptance rates would then require accepting a less-qualified person from one group over a more-qualified person from another, which can itself feel unjust and can hurt the very group it's meant to help by setting people up to fail.
Why it matters: demographic parity is the easiest fairness metric to explain to the public and to regulators, and it directly attacks the worry that a system is excluding a whole group. But it is a blunt instrument. It ignores whether the decision was accurate, and in cases where the groups really do differ on the legitimate target, enforcing it can trade one unfairness for another. It is a useful flag to raise, not a complete definition of justice.
An ad platform shows a high-paying-job ad to 8% of men who match the audience but only 2% of equally-matched women. Demographic parity is violated: the positive outcome (seeing the opportunity) is not delivered at equal rates, regardless of any individual's qualifications.
Demographic parity checks the rate of good outcomes per group — and ignores everything about why.
Demographic parity can be satisfied by a system that is both useless and offensive — e.g. accepting random people from each group at equal rates. Equal rates do not mean equally good decisions. That's why it's usually paired with an accuracy-aware metric like equalized odds.