Counterfactual Fairness

Kusner, Loftus, Russell and Silva introduced counterfactual fairness in a 2017 NeurIPS paper using structural causal models. Informally, a predictor is counterfactually fair for an individual if its prediction in the observed world is the same as its prediction in a counterfactual world in which that individual belonged to a different protected group, given the same background circumstances. The method asks more than whether two groups have equal average outcomes: it asks whether changing the protected attribute in a causal model would change the particular person’s prediction. A structural causal model represents variables and the causal relationships used to generate them. The analyst intervenes on the protected attribute and evaluates the prediction in the resulting counterfactual world. The paper demonstrates the idea in a law-school success prediction example. This does not establish that counterfactual fairness is the uniquely correct ethical or legal standard; it formalizes one fairness intuition under specified causal assumptions. The assumptions carry normative weight. The analyst must decide which variables are causes, which observed features are descendants of protected status, which causal pathways are impermissible, and what “same person in a different group” means. In practice, the relevant counterfactual may be unobservable and the causal graph may be uncertain. Measurement error, hidden confounding and historical discrimination can undermine estimates. A system can satisfy a chosen counterfactual criterion while still producing group-level disparities or harmful outcomes under another fairness definition. Conversely, a predictor can use a protected attribute in a model to correct a harmful proxy path, depending on the causal policy chosen. Reports should publish the model, assumptions, permitted and prohibited pathways, uncertainty and comparisons to other measures rather than presenting one score as proof of fairness.