Reasonable Doubt in the Face of Bias: Fair Flagging with Dirichlet-Based Models

Ensuring safety in machine learning requires not only robustness to adversarial or distributional uncertainty but also protection against systematic bias. Models that produce unfair or group dependent predictions pose critical risks when deployed in socially sensitive domains such as credit, justice, or healthcare. This work introduces EviFair, a fairness-aware safety monitor that flags predictions exhibiting excessive dependence on protected attributes. EviFair combines evidential uncertainty modelling with sensitivity analysis to detect biased decision paths, flagging predictions where fairness cannot be guaranteed. We also show that EviFair’s bias scores can effectively guide post-processing fairness methods. Results on standard fairness benchmark datasets show that EviFair achieves substantial reductions in group disparities with minimal impact on predictive performance, demonstrating its promise as a practical, inference-time mechanism for bias-sensitive, safety-aware model oversight.