A brief safety reminder reduces dangerous clinical decisions in 19 out of 20 AI models

Mount Sinai has analysed more than ten million responses. Potentially harmful choices have fallen from 16.6 per cent to 10.1 per cent, but reminders alone are not enough to ensure safe use.

A study presented by Mount Sinai tests twenty language models against clinical instructions that may compromise safety. The protocol varies the formulations and context, with written scenarios and adapted cases of anonymized output folders.

Out of over 10 million responses, a brief reminder reduces the share of potentially harmful decisions from 16.6% to 10.1%. Nineteen out of 20 models benefit from it. Published in Communications Medicine, the study measures choices in scenarios, without establishing the safety of autonomous use in patients. The authors request renewed testing and clinical supervision.