Will AI build clinical reasoning, or bypass it?

Five new papers, including a Nature Medicine warning about 'never-skilling', suggest that AI's effect on how doctors learn to think depends on supervision and design far more than on the model itself.

Ask a large language model a clinical question and it will hand you an answer that is fluent, confident, and often correct. That is exactly what worries me. Medical education has never really been about producing the right answer on demand. It is about building the reasoning that lets a doctor reach the right answer when the tool is wrong, unavailable, or facing a presentation it has never seen. The anxiety running through this week’s literature is not that AI gives bad answers. It is that good answers, delivered too early and too smoothly, might stop trainees from ever building the reasoning underneath.

I have five papers this week, ranging from a Nature Medicine perspective to a small classroom study, and they circle a single question: does working with AI help students learn to think, or does it let them skip the thinking altogether? The most reassuring finding is that the answer seems to depend far less on the technology than on the conditions we build around it.

Key points:

• A new vocabulary is emerging for the risk: not only the “deskilling” of experienced doctors, but “never-skilling” and “mis-skilling” in trainees who have no baseline to fall back on.

• The best available evidence is genuinely two-sided: supervised, active AI use was linked to stronger critical thinking, while unsupervised exposure to plausible AI errors degraded both accuracy and students’ sense of their own uncertainty.

• Scoping-review evidence leans positive, and AI appears to help less-experienced students the most, but the underlying studies are early and inconsistent.

• We may be measuring the wrong thing, judging AI by accuracy and satisfaction rather than by how it shapes the reasoning process itself.

• Motivation is not the same as depth: an eager AI tutor lifted engagement but did not, on its own, produce higher-order thinking.

Ke et al. set the conceptual frame in a Nature Medicine perspective that gives the problem a name¹. They separate three distinct risks: deskilling, the erosion of competence in experienced clinicians; mis-skilling, the uncritical absorption of flawed AI outputs; and never-skilling, the failure to develop foundational reasoning in the first place because AI has done the cognitive work during the formative years.

The sharpest idea is what they call “false proficiency”, where a trainee performs well in AI-rich clinical settings but collapses in AI-free, high-stakes assessment, because the underlying competence was never built. They are candid that direct evidence in medical training is still absent and present never-skilling as a risk model rather than an established fact. Their proposed answer is a three-phase, competency-protective framework: establish an AI-independent baseline, build critical calibration through structured teaching, then integrate AI under supervision. The practical message is about sequence. Protect the early years, and bring AI in once there is something for it to augment rather than replace.

Ong et al. then show how unsettled the evidence really is, in an editorial for npj Digital Medicine that reconciles two new studies pointing in opposite directions²:

1. In the first, a longitudinal study of 372 senior students using AI diagnostic tools under supervision across a year of rotations, greater AI engagement predicted higher AI literacy, which in turn predicted stronger critical thinking.

2. In the second, a randomised trial of 111 pre-clinical students, even accurate AI explanations produced no benefit, and plausible but wrong ones significantly reduced diagnostic accuracy while leaving students just as confident either way.

The authors argue these results are not contradictory but conditional: supervision and the mode of engagement decide the outcome. Active interrogation of AI within a supervised workflow builds skill, whereas fluent pre-digested answers in an unsupervised test invite what they call cognitive surrender. They also flag a “Matthew effect”, where students with prior technical experience and mastery goals gain the most, which raises real equity concerns.

The reframing I found most useful is their closing one: stop arguing about abstention and start designing the conditions of safe learning, treating this as harm reduction.

Hernández-Rincón et al. widen the lens with a scoping review in the Journal of Investigative Medicine that maps what the empirical literature actually shows³. Across 26 included studies, most reported a positive effect of AI tools on clinical education and decision-making, especially in working through clinical cases and interpreting images and test results.

The most interesting signal is distributional: AI seemed to help students with less clinical experience the most, potentially bringing them closer to the performance of more seasoned peers. The caveats are honest ones. Some studies found no significant difference, imprecise or unclear AI outputs could cause confusion, and one study found that clinical practice guidelines outperformed AI as a support resource.

The authors read this as promise rather than proof, and they call for structured integration into the curriculum rather than the ad hoc use that dominates now. It is a helpful counterweight to the more cautionary pieces, though the youth and inconsistency of the evidence base are impossible to miss.

Zarate makes a quieter but important methodological point in a letter to Medical Teacher⁴. Most studies, he notes, evaluate generative AI using accuracy, efficiency, and user satisfaction. Those measures are necessary, but they say little about how AI shapes the reasoning process itself: how a learner constructs an explanation, generates a differential, or weighs an ambiguous finding. He suggests treating generative AI as a semiotic-epistemic actor, an output that does not merely supply information but foregrounds particular interpretations and implicitly signals what counts as an acceptable clinical answer.

The tension he identifies is that a tool optimised to look right can subtly narrow how students learn to justify decisions and manage uncertainty. His prescription is modest and sensible: study the process, not just the performance, and keep educator guidance in the loop so that AI-assisted reasoning stays reasoning.

Embang et al. bring the theme down to a single classroom in Advances in Health Sciences Education, and their results are instructive precisely because they are mixed⁵. First-year students in a cardiovascular physiology module generated their own questions and used ChatGPT to answer them. Students using the chatbot reported higher autonomy, competence, and task value, although the differences did not reach statistical significance in this small sample.

The revealing part came when the researchers coded the 31 student-generated questions against Bloom’s taxonomy: most sat at “Understand” (41.9%) and “Apply” (45.2%), with very few reaching “Analyse”. In other words, the tool lifted motivation and self-directed engagement, but it did not, by itself, push students toward higher-order thinking. That only happened with instructor mediation to turn information retrieval into genuine dialogic inquiry. The lesson is a useful corrective to the tutor hype: an AI chatbot can be a good scaffold for engagement and a poor substitute for teaching, and the analytical work still needs a human to prompt it.

I wrote about a very similar finding from educational researchers working in UK business schools earlier this year here:

The five papers disagree about plenty, but the throughline is consistent. AI’s effect on clinical reasoning is decided by the conditions we place around it, not by the model. Supervision, sequencing, active interrogation, and attention to process rather than output are what separate a tool that builds reasoning from one that bypasses it. The most useful shift is in the question we ask. We keep asking whether AI helps students learn, when the better question, and the one all five papers gesture toward, is under what conditions AI builds reasoning rather than replacing it. That is a question about curriculum design and teaching, which I find oddly reassuring, because it puts the outcome squarely back in our hands.

1. Ke, Y., Jin, L., Ong, J.C.L. et al. AI-induced never-skilling in medical education. Nat Med32, 1997–2006 (2026). https://doi.org/10.1038/s41591-026-04438-y

2. Ong, A.Y., Sui, M., Rosen, K.L. et al. Reconciling how clinical reasoning is learned in the age of artificial intelligence. npj Digit. Med. 9, 435 (2026). https://doi.org/10.1038/s41746-026-02873-2

3. Hernández-Rincón et al. Effect of the Use of Artificial Intelligence Tools on Clinical Education and Clinical Decision-Making of Medical Students During Clinical Rotations: A Scoping Review. Journal of Investigative Medicine (2026). https://doi.org/10.1177/10815589261460883

4. Zarate, J. J. D. (2026). Beyond performance metrics: Generative AI and clinical reasoning in medical education. Medical Teacher, 1. https://doi.org/10.1080/0142159X.2026.2681964

5. Embang, J.E.G., Tan, Y.H.V., Hooi, S.C. et al. Development of cognitive engagement and motivation using AI chatbot-facilitated questioning in medical education. Adv in Health Sci Educ (2026). https://doi.org/10.1007/s10459-026-10547-7