Group fairness and individual fairness can be incompatible

What it is

Group fairness operates at the statistical level: groups should fare comparably in outcome distributions (selection rates, error rates, score distributions). Individual fairness operates at the case level: similar individuals should be treated similarly regardless of group membership. They look like two phrasings of one ideal but can contradict each other: any mechanism that lifts a group's selection rate to a target adjusts individuals at the margin, producing pairs of applicants with nearly identical profiles and different outcomes that track only group identity. The canonical formulation of individual fairness is "fairness through awareness": define a task-relevant similarity metric and require decisions to be Lipschitz-continuous under it — neighbors receive neighboring outcomes.

Why it happens

The conflict arises because the two fairnesses constrain different objects: group constraints bind outcome distributions, individual constraints bind pairwise cases. When base rates differ across groups, any distribution-aligning mechanism effectively moves some individuals' outcomes by group membership, and those moved individuals are exactly the violations of the individual constraint. Individual fairness carries its own engineering burden: the similarity metric must be task-specific ("similar" for lending means repayment-relevant features, not taste in music), and that definition is itself value-laden and hard to defend — define it badly and individual fairness merely encodes bias into a distance function. Group fairness, in turn, sacrifices within-group heterogeneity: stratifications inside a minority group (recent immigrants versus long-term residents) vanish in group statistics, and individuals at intersections (older women, disabled minorities of a minority) get averaged away at every level of partition.

Studying it

Each fairness can be measured separately: on the group side, report outcome distributions and error rates by group; on the individual side, construct nearest-neighbor pairs (individuals within a similarity threshold under the task metric) and measure their outcome gaps — large gaps violate individual fairness. Studies typically compute both and report the tension: how large are neighbor-pair outcome gaps in a system satisfying its group targets. Methodological cautions: the similarity metric dominates individual-side conclusions, so its definition and robustness must be reported (does the finding flip under an alternative defensible metric?); finer group partitions drain statistical power, and small intersectional cells need uncertainty estimates.

Where it stops holding

Incompatibility is conditional, not universal: with similar base rates, or strong enough model discrimination that case-level rankings are mostly right, both constraints can hold approximately. Individual fairness is most intuitive yet hardest to define in high-stakes case decisions (credit, sentencing recommendations, triage); group fairness suits audit and accountability — external reviewers cannot re-adjudicate every case, so they compare distributions. For products with tiny populations (internal enterprise tools), group statistics are too unstable, and the individual side is the only workable check.

Applying it

• Decide the accountability level first: who reviews this decision and how — external distribution-based audit points to the group side, a case-by-case appeals mechanism points to the individual side; write the choice into requirements rather than leaving it to the ML team's default.
• If individual: define the task-relevant similarity in writing (features, weights, exclusions) and review the definition's value stance; track neighbor-pair outcome gaps as a routine monitoring metric.
• If group: declare the known blind spots for within-group heterogeneity and add intersection-level stratified reports for small groups, preventing "target met, weakest members harmed."
• Verify: output both metric families at least once — a group distribution table plus neighbor-pair outcome gaps; make the tension visible, then justify which side's cost you accept.

Related

• Same group: P4.02.1 Different definitions of fairness conflict mathematically · P4.02.3 Fairness goals must be chosen before design
• Adjacent: P4.09 Algorithmic fairness and disparate impact · P4.03 Data representativeness
• Search terms: group fairness · individual fairness · fairness through awareness