Group vs. Individual Fairness

Group fairness asks whether different populations (like neighborhoods, ethnic groups, or genders) get similar outcomes overall—for instance, whether loan approval rates are roughly equal across communities. Individual fairness asks whether people who are similar in relevant ways get treated the same way, regardless of which group they belong to.

Group fairness measures statistical parity across populations—comparing approval rates, prediction scores, or hiring rates between demographic groups. The core intuition is that if a system works fairly, no protected group should systematically receive worse outcomes. Individual fairness, by contrast, focuses on consistency: people with similar qualifications or circumstances should receive similar decisions, regardless of their demographic identity.

The tension arises because these principles can't be simultaneously maximized. If one group genuinely has higher average qualifications (due to education, experience, or other factors), forcing equal approval rates means either rejecting qualified candidates from the higher-scoring group or lowering standards for the lower-scoring group. Conversely, applying identical decision rules to every individual often produces different group outcomes if the groups differ on relevant factors.

This tradeoff appears everywhere. In hiring, group fairness pushes for diverse representation, while individual fairness insists that the best-qualified candidate gets the job regardless of demographics. In lending, group fairness flags redlining patterns, while individual fairness requires that creditworthy applicants get approved regardless of neighborhood. In healthcare, group fairness audits whether treatment recommendations vary by race, while individual fairness ensures that patients with identical symptoms and history receive identical care. The best approach depends on context: group fairness excels at detecting systemic bias, while individual fairness protects meritocracy and prevents reverse discrimination.

Fairness in machine learning requires contextual value judgments about which principles matter most for specific applications. Different fairness criteria often involve trade-offs rather than pure technical optimization.

The framework tends to treat group membership as fixed and observable, missing intersectional considerations (how race, gender, and class interact) and situations where group identity is sensitive, self-identified, or fluid. Finally, formal fairness definitions focus on statistical properties at the expense of procedural fairness—whether the decision-making process itself is transparent, contestable, and perceived as legitimate by affected communities.