Deskilling, Never-Skilling, Mis-Skilling: 3 Risks Reshaping Medical Education

Deskilling, never-skilling, and mis-skilling are three distinct risks that artificial intelligence poses to physician training, named together for the first time in a 2025 New England Journal of Medicine framework.

Deskilling is the erosion of a skill a physician already has. Never-skilling is the failure to ever develop a foundational skill because AI performed the task first. Mis-skilling is the quiet adoption of an AI tool's errors as one's own clinical reasoning. A 2026 framework in Nature Medicine has since sharpened the picture, warning that never-skilling produces "false proficiency", competence that looks real until the AI is taken away.

Where did this framework come from?

In August 2025, NEJM published "Educational Strategies for Clinical Supervision of Artificial Intelligence Use" by Abdulnour, Gin, and Boscardin. As an assistant dean, a practicing neonatologist, and someone who has spent years at the intersection of EdTech and health professions education, I've watched AI move from a curiosity to a bedside tool faster than any technology in my career. This paper was the first to give that shift a precise vocabulary.

The core argument: we are training a generation of physicians who may never have to struggle through a differential diagnosis unassisted. Its in that exact struggle, turning over the data, sitting with uncertainty, being wrong and correcting course, is where clinical reasoning is actually built. Remove the struggle, and you risk removing the skill.

What is deskilling, exactly?

Deskilling describes the erosion of a competency a clinician already possesses. An experienced radiologist who leans on an AI flagging tool for years may gradually lose some of the pattern-recognition sharpness that took a decade to build. This isn't a new fear in medicine. In the past clinicians worried aloud about the stethoscope, the calculator, and the electronic health record in much the same terms. A 2025 NEJM AI historical analysis, "Cognitive Aids, Artificial Intelligence, and Deskilling in Medicine," traces this anxiety back more than a century and finds that every major cognitive aid has triggered the same debate. The authors' point isn't that the anxiety is wrong, it's that AI's scale and speed make the stakes higher than any prior tool.

What is never-skilling — and why a new study calls it the most dangerous of the three?

Never-skilling is different from deskilling because there's no prior competency to lose. A trainee who routinely lets a large language model generate the first-pass differential may never build the underlying reasoning skill at all, because the AI did the cognitive work during the exact developmental window when that skill should have formed.

A 2026 Nature Medicine Perspective, "AI-induced never-skilling in medical education," goes further than the original NEJM framework and gives this risk a name that should worry every program director: false proficiency. The authors describe it as competence that holds only in the presence of AI and collapses without it. As they put it, the trainee "logs the hours but bypasses the mental work those hours are meant to provide", and that gap follows them into early independent practice, often invisibly, because their AI-assisted performance looks indistinguishable from genuine competency right up until the AI isn't there.

This tracks with what a wave of 2026 research is now finding in practice. Coverage of a recent study on AI overuse among trainees described measurable declines in independent critical thinking tied to heavy, unscaffolded AI use, not because trainees became less capable, but because the reasoning muscle was never exercised in the first place.

What is mis-skilling?

Mis-skilling is the most insidious of the three because it doesn't feel like a deficit, it feels like learning. It occurs when a trainee uncritically accepts an AI's error and internalizes it as fact, building future clinical judgment on a flawed foundation. Unlike deskilling (losing something real) or never-skilling (never gaining something), mis-skilling actively replaces correct knowledge with incorrect knowledge, delivered with the same fluent, confident tone AI tools use for everything else they generate.

What is the DEFT-AI framework?

The NEJM authors propose a supervisory model called DEFT-AI, designed to help clinical educators guide trainees through AI interactions the way we've long guided them through patient encounters. The framework asks supervisors to engage trainees in active Discussion of the AI's output, weigh the Evidence behind it, give direct Feedback on how it was used, explicitly Teach the underlying reasoning the AI shortcut past, and build a habit of Recommendation, deciding, case by case, when AI assistance is appropriate at all.

The point of DEFT-AI isn't to ban AI from training. It's to insert a human checkpoint at exactly the moment AI tends to remove one.

What should medical schools do now?

Most programs are not ready for this conversation. Many are still debating whether to allow AI tools in the library while their residents are already using them at the bedside. The institutions making real progress, and we'll dig into specific examples in the next post in this series, are treating AI literacy as a vertical thread through training, not a single elective bolted onto year two.

We have a choice. Either medical education leads the integration of AI, with deliberate checkpoints like DEFT-AI built in from day one, or AI integrates itself into medicine while we clean up the mess afterward. The first path is harder. It's also the only one that protects the next generation of patients.

FREQUENTLY ASKED QUESTIONS

Q: What is DEFT-AI?

A: DEFT-AI is a supervisory framework proposed in a 2025 NEJM paper for guiding trainees through clinical AI use. It stands for Discussion, Evidence, Feedback, Teaching, and Recommendation. These five checkpoints are designed to keep a human reasoning step in the loop whenever AI is used in training.

Q: What is "false proficiency" in medical training?

A: False proficiency, a term introduced in a 2026 Nature Medicine Perspective on never-skilling, describes AI-assisted performance that appears competent but collapses once the AI tool is removed, because the trainee never built the underlying independent reasoning skill.

Q: Is AI making doctors worse?

A: Not inherently. The risk is concentrated in how AI is used during training, not whether it's used at all. The NEJM and Nature Medicine frameworks both argue that AI can strengthen physician training when paired with deliberate human checkpoints, and can quietly erode it when used as an unsupervised shortcut.

Q: How can residency programs guard against never-skilling?

A: Early evidence points to structured approaches: requiring trainees to generate an independent differential before consulting AI, grading the reasoning process rather than only the final answer, and building in AI-free assessment moments, supervised practical exams and direct observation, that can't be shortcut.