Bias Detection and Mitigation in Training Data for Argumentation-Based Legal AI

Investigate the presence of biases in training data used to develop argumentation-based explainable AI models for legal decision-making, and derive methods to remove or mitigate such biases to improve precision and applicability in judicial contexts.

Background

The paper argues for argumentation-based explainability as the most robust approach for legal AI, but it emphasizes potential risks from biased training data that can undermine fairness, legality, and trust. The authors explicitly state that further exploration remains open on identifying and mitigating such biases.

Addressing this problem is central to aligning argumentation-based systems with evolving European regulatory standards (GDPR and AIA) and ensuring their deployment in high-risk legal settings where decisions must be transparent and contestable.