Fairness Implications of Heterogeneous Treatment Effect Estimation with Machine Learning Methods in Policy-making

Standard fairness tools designed for automated decision-making may not work for policy applications: The paper argues that common AI fairness methods were built for systems that make decisions directly, but government policy-making typically uses machine learning models to inform human decision-makers rather than replace them—a crucial distinction that changes how fairness should be approached.

Complexity of causal machine learning models can obscure important information from policymakers: The paper identifies that while causal machine learning methods are useful for understanding how policies affect different groups differently, their complexity can make it difficult for human decision-makers to understand the results well enough to make fair policy choices based on them.

Fairness in policy contexts requires models that enable accurate human judgment, not just unbiased predictions: Rather than adjusting models to remove bias (the traditional fairness approach), the paper proposes that fairness in policy should focus on whether the model provides decision-makers with the information they need to make value judgments about just outcomes—a fundamentally different goal.

The indirect nature of government decision-making requires rethinking fairness safeguards: Because policymakers retain ultimate decision-making authority when using these models, fairness protections should focus on ensuring transparency and interpretability rather than constraining model outputs the way they would for fully automated systems.