According to a report by Futurism today (the 12th), Simar Bajaj, a medical student and journalist at Stanford School of Medicine, and Joseph Sakran, a trauma surgeon and executive vice chair of surgery at Johns Hopkins Medical Center, believe the problems caused by medical students using AI are already deeply concerning.
Bajaj and Sakran wrote in The Guardian that the dangers of medical students using AI involve not only skill deterioration, but also the possibility that they may never truly develop those skills in the first place. “A doctor who has forgotten how to reason may still be able to relearn the skill, but a doctor who never learned how to reason may not be able to do so.”
Bajaj and Sakran pointed to two recent findings that are particularly concerning.
The first comes from internal data. About two-thirds of doctors in the United States are already using OpenEvidence, an AI chatbot designed specifically for clinicians, indicating that the technology has already entered medical workflows on a broad scale and become deeply integrated into healthcare professionals’ daily work.
The second is a new study published in Nature Medicine, in which researchers evaluated the reliability of medical large language models. The results were surprising: these clinical AI tools performed worse than general-purpose chatbots such as ChatGPT and Claude when answering medical questions.
The study authors said their overall performance was roughly comparable to Google’s “AI Overviews.” Previously, AI Overviews had become “infamous” for its inconsistent performance and inaccurate answers.
Even if AI tools can help locate information—notwithstanding the major reliability issues that remain—they cannot replace learners’ own thinking. The reason is that, during the learning process, the difficult thinking itself is the point.
Bajaj and Sakran wrote: “For example, when trainees were once asked to list possible diagnoses, they might have had to struggle through the problem and ultimately produce an incomplete list, learning from what they missed and sometimes feeling embarrassed as a result. Now, trainees need only ask OpenEvidence to receive an almost perfect answer, including possibilities they had never considered, without having to experience the embarrassment of missing a diagnosis.”
Bajaj and Sakran wrote that AI differs from previous advances in medical technology because it is not merely expanding what doctors can see; it is intervening in the cognitive mechanisms that medical training is supposed to build.
Bajaj and Sakran proposed several solutions. One would require trainees to handle cases independently on a regular basis with no AI assistance, similar to the U.S. Federal Aviation Administration’s recommendation that pilots periodically turn off the autopilot and fly the aircraft manually. Another powerful constraint might come from “shame.”