Study finds AI safety refusals causing medical harm by withholding critical health guidance

A pre-registered arXiv paper published on April 9, 2026, by researcher David Gringras introduces 'IatroBench,' a benchmark designed to measure harm caused by AI over-refusal in medical contexts. The study analyzed 3,600 model responses across 60 clinical scenarios, scoring them on both commission harm — when AI says something unsafe — and omission harm — when AI withholds genuinely needed information. Gringras argues that current AI safety guardrails, calibrated almost entirely to prevent dangerous outputs, are systematically failing patients who seek legitimate medical guidance, such as those managing prescription tapers without a doctor. The paper contends this represents an iatrogenic problem, where the safety apparatus itself inflicts harm on the people it was built to protect. The research has drawn significant attention in online communities at the intersection of chronic illness and machine learning since its release.