Counterfactual Fairness

Unlike observational metrics, counterfactual fairness is defined using a Structural Causal Model (SCM) . This requires explicitly mapping the causal relationships between protected attributes, mediators, confounders, and outcomes.

• Nodes represent variables (e.g., education, income).
• Edges represent direct causal links.
• The model computes a counterfactual prediction by intervening on the protected attribute node while holding exogenous noise factors constant.

This metric provides a guarantee at the level of a single person, not just group averages. It asks: 'Would this specific applicant have been approved if their race/gender were different, but all other causally independent attributes remained the same?'

• Contrasts with Demographic Parity, which only equalizes acceptance rates across groups.
• Contrasts with Equalized Odds, which only balances error rates across groups.
• A model can satisfy group fairness metrics while still being counterfactually unfair to specific individuals.

Counterfactual fairness allows for path-specific fairness analysis. Not all causal paths from a protected attribute to an outcome are unfair.

• Unfair Path: Race → Zip Code → Loan Decision (redlining proxy).
• Fair Path: Race → Genetic Predisposition → Medical Diagnosis (biological causality). The framework allows developers to block unfair causal pathways while preserving legitimate, explanatory ones, avoiding the 'leveling down' problem of simpler metrics.

Counterfactual fairness resolves the infra-marginality problem that plagues observational definitions. Observational metrics cannot distinguish between two individuals who are observationally identical but differ in their counterfactual outcomes.

• Example: A high-achieving student from a disadvantaged background might have a very different counterfactual outcome than a privileged peer with the same observed score.
• By modeling the latent noise variables (U), the SCM captures this unobserved heterogeneity, providing a more philosophically sound notion of equity.

Deploying counterfactual fairness requires strong assumptions and specific technical inputs:

• Causal Graph Specification: Requires domain expertise to define the correct DAG.
• Structural Equations: Requires modeling the functional form of causal relationships.
• Latent Variable Inference: Requires inferring the posterior distribution of unobserved noise variables (U) from observed data.
• Sensitivity Analysis: Because the causal model is often uncertain, rigorous sensitivity analysis is required to test how violations of assumptions affect the fairness guarantee.

Counterfactual fairness aligns closely with the right to explanation and contestability requirements in high-risk AI governance.

• It provides a natural language explanation: 'Your loan was denied because of factor X. If your demographic attribute had been different, the decision would/would not have changed.'
• This directly supports the Algorithmic Impact Assessment process by providing a quantitative measure of the causal influence of protected attributes.
• It helps demonstrate that a system is not a Solely Automated Decision based on a prohibited proxy.