Green Shielding: A User-Centered Approach for Trustworthy AI

The paper proposes Green Shielding, a user-centered research agenda that, by characterizing how benign input changes alter model behavior, builds evidence-backed deployment guidance for high-risk settings and operationalizes it with the CUE standard. Guided by the PCS framework, the research team worked with practicing physicians to build the HealthCareMagic-Diagnosis (HCM-Dx) benchmark in medical diagnosis based on real patient questions, together with a structured reference diagnosis set and clinical metrics oriented toward differential diagnosis lists. Experiments on multiple frontier large language models show that prompt-level factors shift model behavior along clinically meaningful dimensions, forming Pareto-like trade-offs; among these, neutralization improves plausibility, makes differential diagnoses more concise and closer to clinician style, while reducing coverage of high-probability and safety-critical diseases. The authors state that these results show that interaction choices can systematically change task-relevant properties of model outputs and can support safer deployment in high-risk domains, and that the agenda can be extended to other decision-support settings and agent systems.