The Stability Dilemma: Addressing Uncertainty in Public Sector AI

Increasing adoption of AI in public sector decision-making presents significant challenges due to the inherent uncertainties of human behavior and volatile environments. It has been suggested in prior literature that effectiveness of AI is predicted by the stability and predictability of the environment, viz., the Adapt-to-AI principle. Unlike domains such as transportation where stability can be enforced to optimize AI performance through establishing protocols, public sector domains like law, healthcare, and welfare require flexibility, interpretability, and procedural fairness. We examine how uncertainty affects AI effectiveness in the public sector, and consider whether human behaviors and environments should be adapted to improve AI utility. Using the Adapt-to-AI principle as a foundation, we propose a structured framework for analyzing AI uncertainty factors, including the intended AI user, AI subject, physical environment, and social environment. Focusing on judicial AI, we assess the feasibility and ethical implications of adapting public institutions to accommodate AI systems effectively. While some adaptations are highly feasible, adapting human behaviors and legal frameworks introduces significant ethical risks. We argue adaptations should be made to integrate AI systems from the ground up-but as a decision-support tool rather than a determinative system, ensuring human oversight remains central to public governance. Finally, we outline future research directions on adaptation strategies and governance frameworks, promoting ethical and effective AI deployment.