nature.com web signal

Medical AI Misinformation Harms Trainees More Than AI Helps

TL;DR

  • A randomized trial of 111 students found misleading AI explanations significantly degraded diagnostic accuracy while correct AI explanations showed no measurable improvement.
  • A longitudinal study tracked 372 senior medical students over 12 months using AI tools in supervised clinical rotations.
  • Authors argue the findings have direct implications for how medical curricula should be designed and clinical competence defined.

The uncomfortable question in AI-augmented medical education has always been whether AI tools help students learn to reason or merely give them answers to copy. An article in npj Digital Medicine offers the sharpest empirical answer yet, and it runs asymmetric.

Drawing on two recent studies, the piece examines what happens when students actually use AI tools in clinical training. One study, a longitudinal survey of 372 senior medical students across 12 months of clinical rotations, tracked self-reported outcomes as students used AI tools, including computer-aided radiology systems and EHR-based clinical decision support, under the supervision of senior clinicians. The other was a randomized trial with 111 students designed to isolate whether AI explanations help or harm diagnostic reasoning. What it found: misleading AI explanations significantly degraded diagnostic accuracy, while correct AI explanations offered no significant improvement over a no-explanation control. The authors describe the result as a "significant and problematic asymmetry."

That asymmetry matters because it reframes the standard debate. The question is not just whether AI helps students learn. It is whether the harm profile when AI is wrong in a plausible-sounding way outweighs any learning benefit. Here, the answer is yes: "the benefits of correct AI explanations do not outweigh the risks of plausible misinformation."

The honest caveat the article itself supplies is that "the effects of AI may depend on the individual and educational environment." What the reporting does not give you is clarity on whether these harms persist as trainees gain more clinical experience, or whether the results generalize beyond the radiology and EHR-focused settings studied here.

For curriculum designers and medical AI developers alike, the practical direction the findings point toward is structured, supervised exposure over open-ended AI access, and treating plausible error reduction as at least as important an engineering goal as raw diagnostic accuracy.