Counterfactually Fair Regression with Double Machine Learning

The method uses machine learning to isolate discrimination: The paper proposes Double Machine Learning (DML) Fairness, which uses machine learning to separate out the effects of sensitive characteristics (like race or gender) from other factors, allowing decisions to be based on what outcomes individuals would have achieved without discrimination.

Fairness is measured at the individual level: Unlike some fairness approaches that focus on group-level statistics, this method aims to make individual-level outcomes "counterfactually fair"—meaning each person is evaluated as if they didn't have their sensitive characteristic, which the authors demonstrate in hiring and law school GPA applications.

The approach has important real-world limitations: The paper acknowledges that DML Fairness assumes effects of variables are "additively separable" (can be cleanly separated mathematically), which may not hold when real-world discrimination is complex and interconnected—raising questions about when this technical solution actually addresses real discrimination.

Achieving fairness requires more than technical fixes: The authors explicitly discuss whether their method "can achieve justice in these settings," suggesting that technical fairness approaches alone may be insufficient for addressing systemic discrimination where social concepts are "conceptually complex."

Causal reasoning improves fairness estimation: By drawing from causal inference methods (the Potential Outcomes framework), the approach moves beyond correlation-based fairness to consider what would have happened under different circumstances, providing a more principled foundation for counterfactual fairness.