Individual fairness

Individual fairness is a criterion of algorithmic fairness proposed by Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel in 2012. It requires that similar individuals receive similar treatment: if two individuals are alike in all relevant respects, the decision system should produce similar outcomes for them.

The criterion is motivated by an intuitive moral principle: fairness operates at the level of individuals, not groups. A system that treats similar people differently is unfair, regardless of whether the outcomes are balanced across demographic categories. Individual fairness captures the anti-discrimination intuition that what matters is how people are treated, not which group they belong to.

The difficulty is operationalizing "similarity." The similarity metric must be task-specific and is itself a design choice with profound consequences. A poorly chosen metric can encode historical biases into the fairness criterion itself. If the similarity function treats zip code as a legitimate feature, it may proxy for race. If it treats credit history as neutral, it may encode decades of discriminatory lending.

Individual fairness and demographic parity are in tension. A system can satisfy individual fairness while violating demographic parity if the population distribution produces unequal outcomes for equally qualified individuals. Conversely, a system can satisfy demographic parity while violating individual fairness by forcing parity at the expense of treating similar individuals differently. The impossibility of fairness results show that no system can satisfy all criteria simultaneously.

The deeper critique is that individual fairness assumes the existence of a neutral similarity metric, but in socially structured domains, no such metric exists. What counts as "relevantly similar" is itself contested. The algorithm does not resolve the contest. It chooses one side.