CliniCARE-Bench: Clinical AI Agents Can Be Right for the Wrong Reasons

CliniCARE-Bench is a new benchmark for evaluating clinical AI agents, which tested 16 systems on 750 real patient cases. The key finding is that systems often arrive at correct answers for the wrong reasons, taking prohibited shortcuts or failing to abstain when evidence is insufficient. When verdicts are only credited if the investigation process is sound, accuracy scores drop significantly and reorder the model leaderboard. This research emphasizes that for high-stakes applications like healthcare, evaluating the agent's reasoning process is more critical than just the final outcome. The benchmark provides a framework for auditable AI by scoring the investigation, evidence grounding, and policy use.