Nature maps the risks of large language models in hospitals, in three families

Security, model reliability, and human interaction: a review published on August 19 proposes a unique framework and recommends dedicated AI security operations centers.

A review published in Nature August 19, 2026 (DOI 10.1038/s41586-026-10687-1), led by Jan Clusmann at the Else Kröner Fresenius Centre for Digital Health at the Technical University of Dresden, draws up a structured inventory of the risks posed by major language models in clinical practice.

The authors group these risks into three families. Security threats first: data poisoning, quick injections, infrastructure vulnerabilities that can expose patient information. The reliability of the model then, with plausible but false outputs that create a direct diagnostic hazard. Finally, human interaction problems, where automation bias and confirmation bias distort clinical decision-making.

The proposed safeguards

The review recommends a secure development with rigorous training data maintenance, continuous performance monitoring, institutional oversight teams, operational security centres (SOCs) capable of detecting AI-specific incidents, and regulatory frameworks adapted to systems that evolve after they are put into operation. "User safety is not self-evident," says Jan Clusmann: risks appear at many stages and need to be addressed systematically, human supervision remains indispensable.