Medical Training Faces 'Never-Skilling' Risk as AI Tools Reshape Learning

Medical educators are confronting a new challenge that goes beyond doctors losing skills to automation: trainees who may never develop clinical reasoning abilities in the first place.

The concern centers on tools like OpenEvidence, an AI chatbot for clinicians now used by approximately two-thirds of US doctors. While experienced physicians can critically evaluate AI-generated answers, medical students and residents are incorporating these tools at a formative stage—before they've built independent clinical judgment.

Simar Bajaj, a medical student at Stanford, and Dr. Joseph Sakran, a trauma surgeon at Johns Hopkins Medicine, argue that this creates a "never-skilling" problem distinct from traditional deskilling. A doctor who has forgotten how to reason can potentially relearn, they note, but one who never learned may not recover those abilities.

The apprenticeship model at risk

Medical training has historically functioned as an apprenticeship where students learn through struggle and repetition under supervision. A trainee asked to generate a differential diagnosis might offer an incomplete list, then learn from the gaps—sometimes painfully. Now that same trainee can query OpenEvidence for a comprehensive answer within seconds, appearing prepared while concealing the deficit that training should reveal.

The authors point to a recent Nature Medicine study showing that AI tools pulling from medical literature can be less reliable than they appear and sometimes less accurate than general-purpose chatbots. Yet trainees report feeling trapped in an arms race: if peers use AI to sound more prepared, opting out feels like a competitive disadvantage.

Why it matters

As AI becomes standard in clinical practice, the medical profession must ensure the next generation can supervise these tools effectively. Physicians trained alongside AI may struggle to question reasoning that shaped their own understanding. This creates a potential blind spot precisely when critical oversight becomes most important—a generation of AI supervisors who were never independent reasoners first.

Proposed safeguards

Bajaj and Sakran recommend structural changes rather than relying on individual restraint. Medical schools should require trainees to "reason first, consult AI second," making unaided assessments visible before turning to automated tools. A resident admitting a patient overnight might write a brief pre-AI assessment; on rounds, attendings could pause before anyone consults AI to discuss how new information changes the diagnosis.

They draw parallels to aviation, where the Federal Aviation Administration advises pilots to maintain manual flying skills by periodically disengaging autopilot. Medicine needs similar discipline, with trainees required to work through no-AI cases and assessed on unaided reasoning.

The authors also propose "flight simulator" drills using real clinical cases with AI-generated assessments containing subtle flaws. Trainees would learn not reflexive skepticism but disciplined judgment—when to trust tools, when to question them, and when to catch errors.

The goal isn't making training harder for its own sake, the authors emphasize. A trainee who has seen pneumonia that looks like heart failure, then heart failure that looks like pneumonia, develops richer bedside judgment than any AI can provide: knowing what to notice, what to question, and when familiar patterns should be distrusted.

These details were first reported by Simar Bajaj and Joseph Sakran in The Guardian.