A new medical AI study found the same flaw in OpenEvidence, OpenAI, Anthropic, and Doximity

Fortune · 18

A Stanford-led benchmark called NOHARM tested AI systems from OpenEvidence, OpenAI, Anthropic, and Doximity on 1,100 real clinical cases and found a common flaw: all models frequently omit important information rather than stating falsehoods, with 76.6% of harmful errors being omissions. Doximity's Ask tool performed best in the study, though OpenEvidence disputed the methodology. The findings highlight that current medical AI systems maintain what cardiologist Eric Topol calls an "illusion of readiness," creating liability questions as regulators and hospitals decide who bears responsibility when AI suggestions prove wrong.