Study finds AI reasoning models trade safety for diagnostic accuracy
Ongoing: AI Cooling
A study in the journal Nature Communications tested large reasoning AI models on clinical diagnosis. It reports that the models diagnose more accurately but raise more false alarms.
The study also found that the models give less consistent outputs than before. It reports that oversight by physicians keeps the gains in accuracy while lowering the safety risks. The research is peer-reviewed. Only a short summary of it was available to us, so the sample size, the datasets and the methods are not known. The full paper is needed to judge how strong the findings are.