Unmasking AI Explainable Models for Bias Detection in Criminal Justice

This research paper, Unmasking AI: Explainable Models for Bias Detection in Criminal Justice, explains how Explainable AI (XAI) techniques, particularly LIME and SHAP, can be used to increase transparency and lessen bias in criminal justice risk assessment tools, like recidivism prediction systems. Through the creation of a prototype that blends XAI with fairness-aware machine learning techniques, such as adversarial debiasing, the work addresses the pressing issues of black-box AI opacity and systemic bias. The work measures algorithmic bias using metrics such as disparate impact and equal opportunity difference, utilizing publicly available datasets like COMPAS, and demonstrates quantifiable improvements (e.g., reducing divergent impact from 0.68 to 0.89 after XAI tuning). The framework, which is intended for judicial stakeholders, complies with ethical norms and policy criteria such as the EU AI Act and offers visually understandable explanations (such as feature attribution plots). This project aims to bridge the gap between high-accuracy AI (76% precision) and fair outcomes in critical legal decision-making by making tools open-source and placing a strong emphasis on replication.