Guardian op-ed: AI risks 'never-skilling' medical trainees

An opinion piece in the Guardian by Simar Bajaj, a Stanford MD candidate and Knight-Hennessy Scholar, and Joseph Sakran, a Baltimore trauma surgeon, argues that medical trainees leaning on AI tools before they have built their own diagnostic reasoning may never develop the judgment safe practice actually requires. Two AI experts in AI Weekly's Who's Who directory circulated the piece within its first day online.

The concern lines up with a term introduced earlier in 2026 in a Nature Medicine perspective: never-skilling. Where deskilling describes experienced clinicians losing capabilities through disuse, and mis-skilling describes trainees uncritically absorbing AI errors as fact, never-skilling names a distinct problem. A student who logs the hours but bypasses the mental work never lays down the foundational competence in the first place. That framework called the result 'false proficiency': performance that looks real until the AI is taken away. The three-risk taxonomy itself was first named together in a 2025 New England Journal of Medicine paper by Abdulnour, Gin, and Boscardin.

Underneath the argument is the older idea of productive struggle. Working through a differential slowly, being wrong, and noticing why, is what teaches a clinician to catch an error later, including one the model just made. If the tool is present at every step of that struggle, the error-detection instinct never gets built.

Two limits on the case are worth naming. The Nature Medicine framework itself concedes that direct evidence from medical training is absent; the argument is grounded in learning theory and non-clinical studies, not randomized data on doctors. And the op-ed does not identify which medical schools are already changing curricula in response, or who funds the extra faculty time an AI-aware supervision model would require.

The constructive read is that the window to act is now, before AI-native cohorts start rotating on wards. Building AI-off periods into pre-clinical years, running simulations against deliberately flawed AI outputs, and pairing every AI-assisted encounter with structured attending oversight are tractable curriculum changes, and cheaper than trying to reverse a class of graduates who never built the reflex to catch a wrong answer.