Fair Recourse for All: Ensuring Individual and Group Fairness in Counterfactual Explanations

Counterfactual explanations can perpetuate unfairness: The paper identifies that standard counterfactual explanations—which show people how to change their circumstances to get a different decision from an AI system—may offer different or less actionable suggestions to similar individuals or to people from different demographic groups, undermining fairness in AI decision-making.

The paper proposes a unified fairness framework: Rather than treating individual fairness (similar people get similar suggestions) and group fairness (different demographic groups get equitable suggestions) as separate concerns, the authors develop a single approach that addresses both simultaneously through reinforcement learning, showing these objectives don't have to be in conflict.

Fairness in explanations requires new metrics: The research extends existing fairness auditing metrics to evaluate counterfactual explanations specifically, introducing measures like "equal choice of recourse" and "equal effectiveness" to assess whether the actionable suggestions an AI system provides are truly fair across individuals and groups.

The solution is model-agnostic: The proposed approach works with any machine learning model without requiring access to its internal workings, making it practical for real-world deployment across different AI systems used in high-stakes decisions like lending or hiring.

Fairness in recourse matters for trust: By ensuring that people receive fair and actionable explanations for AI decisions affecting them, this work addresses a key component of trustworthy AI—not just explaining why a decision was made, but offering equitable paths forward for those affected by it.