The DOI matches its Crossref record; 42 of the 42 references it cites were found in a registry or an index; 20 of the 49 sources its claims cite were checked for support.
AI-induced never-skilling in medical education
Open accessAbstract
The integration of artificial intelligence (AI) into medical training is accelerating faster than the educational frameworks designed to govern it. This Perspective identifies a risk that has received insufficient attention: that trainees who rely on AI during the early formative years of clinical education may fail to develop the foundational reasoning skills that safe, independent practice requires. We refer to this as 'never-skilling', distinguishing it from deskilling in experienced clinicians and from mis-skilling, in which uncritical acceptance of AI errors leads trainees to internalize flawed clinical knowledge as fact. Although direct evidence from medical training is absent, the concern is grounded in established learning theory and supported by early empirical signaling from nonclinical settings. AI is not inherently harmful to learning; its educational impact depends on how and when it is introduced. We propose a three-phase competency-protective framework: establishing AI-independent baseline competency, building critical calibration through structured pedagogy, and integrating AI under supervision in medical training. This is a pedagogy research agenda that requires further empirical investigation to ultimately inform future policy recommendations.
Overview
CoChat Summary
## TL;DR
This Perspective argues that AI use during formative medical training could cause “never-skilling”: trainees may appear competent with AI but fail to develop independent clinical reasoning. The paper proposes a three-phase, competency-protective framework and research agenda, but emphasizes that direct causal evidence in medical trainees is not yet available.
## Key findings
- The paper defines never-skilling as the “failure to develop foundational competencies during formative training due to excessive AI substitution for cognitive effort,” which can prevent AI-independent competence needed for progression, graduation, or licensure.
- It distinguishes never-skilling from deskilling and mis-skilling; in Figure 1, it describes never-skilling as a period of “false proficiency” in which AI-assisted performance rises above the competency threshold but performance returns toward novice level when AI is withdrawn.
- The authors note that direct causal evidence in medical trainees is absent, but cite indirect signals including a 6.0% reduction in adenoma detection rates during unassisted colonoscopies after routine AI-supported practice and a study in high school students where AI tutor access led to approximately 17% lower outcomes on subsequent unaided assessments.
- In the evidence base, the paper also cites a survey of 666 adults finding a negative association between AI tool use and critical thinking, and a meta-analysis of 106 experimental studies reporting that human-AI combinations frequently underperformed the best individual decision-maker on clinical decision tasks.
## Methods
- This is a Perspective paper, not an original empirical study; it synthesizes learning theory, prior literature, and early empirical signals to develop a conceptual risk model and framework.
- The paper proposes a three-phase framework: Phase 1 AI-free foundational competency, Phase 2 guided integration with adversarial calibration, and Phase 3 supervised AI collaboration in residency.
- It also outlines a research agenda calling for prospective cohort studies and a longitudinal randomized study comparing levels and types of AI support, with AI-independent competency outcomes assessed at graduation and follow-up; the paper does not report having conducted these studies.
## Limitations
- The paper explicitly states that direct causal evidence linking AI exposure during training to competency failure does not exist, that the evidence cited is indirect and predominantly non-clinical, and that never-skilling is a risk model, not an established phenomenon.
- It states that the prevalence, severity, and reversibility of AI-induced competency deficits are unknown, and that properly scaffolded AI integration may instead enhance foundational skill development.
- The framework is described as untested; its effectiveness, feasibility, and unintended consequences have not been evaluated, and the paper recommends pilot implementations before widespread adoption.
- Evident from the text: much of the argument is theoretical and hypothesis-driven, so its claims about medical trainees remain unconfirmed; also, some cited supporting evidence comes from non-clinical settings, limiting direct applicability.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
15 Valid · 13 Not valid · 28 Not checked · 122 Own results and interpretations
At least one cited source doesn’t support this claim or points to a different paper, and none confirmed it — see the verdicts below.
In the paper
“well-structured AI learning tools have been shown to improve examination performance 10”
What the sources say
- Disputedcites [10] Leong, Y. H., Nambiar, L., Tay, V. Y. J., Lie, S. A. & Yuhe, K. Feasibility of a specialized large language model for postgraduate medical examination preparation: single-center proof-of-concept study. JMIR Form. Res. 9, e77580 (2025). sourceExperimental — accuracy not yet measured
The Discussion’s Limitations and Future Work section states that the study did not examine whether PASS affected actual examination performance or learning outcomes. It reports user acceptance and grading agreement instead, so it does not show improved examination performance. (LLM, conf=0.88)
Checked by a reasoning model — may be wrong.
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Disputed
At least one cited source doesn’t support this claim or points to a different paper, and none confirmed it — see the verdicts below.
In the paper
“Leong et al. demonstrated that AI-assisted learning improved postgraduate examination performance 10 .”
What the sources say
- Disputedcites [10] Leong, Y. H., Nambiar, L., Tay, V. Y. J., Lie, S. A. & Yuhe, K. Feasibility of a specialized large language model for postgraduate medical examination preparation: single-center proof-of-concept study. JMIR Form. Res. 9, e77580 (2025). sourceExperimental — accuracy not yet measured
The Limitations and Future Work section explicitly says the study did not examine whether using PASS affected actual examination performance or learning outcomes. Its Conclusions report feasibility and user acceptance, not demonstrated improvement in examination performance; page numbers are not provided in our copy. (LLM, conf=0.90)
Checked by a reasoning model — may be wrong.
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Disputed
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Large language models (LLMs), computer vision systems, and clinical decision-support systems (CDSS) have demonstrated substantial clinical value across medical specialties 1 -4 .”
What the sources say
- Partly supportedcites [1] Bajwa, J., Munir, U., Nori, A. & Williams, B. Artificial intelligence in healthcare: transforming the practice of medicine. Future Healthc. J. 8, e188–e194 (2021). sourceExperimental — accuracy not yet measured
The Diagnostic imaging section reports strong diagnostic performance for image-based AI across several specialties, supporting the computer-vision portion. The paper does not report demonstrated clinical value for large language models or clinical decision-support systems; its Evaluate and validate section cautions that high model performance alone does not establish clinical utility. Page numbers are unavailable in our copy. (LLM, conf=0.88)
- Not checked — no full text of the sourcecites [2] Teo, Z. L. et al. Generative artificial intelligence in medicine. Nat. Med. 31, 3270–3282 (2025). source
- Not checked — no full text of the sourcecites [3] Rajpurkar, P., Chen, E., Banerjee, O. & Topol, E. J. AI in health and medicine. Nat. Med. 28, 31–38 (2022). source
- Partly supportedcites [4] Wu, J. et al. Vision-language foundation model for 3D medical imaging. NPJ Artif. Intell. 1, 17 (2025). sourceExperimental — accuracy not yet measured
The Abstract describes potential clinical benefits of vision-language models in medical imaging, and ‘Addressing computational and clinical challenges in 3D’ identifies possible supportive uses. But that section says prospective trials are still needed to demonstrate clinical value. The paper does not establish that LLMs, computer vision systems, and CDSS have already demonstrated substantial clinical value across medical specialties; page numbers were not available in our copy. (LLM, conf=0.84)
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Vaccaro et al .'s meta-analysis of 106 experimental studies found that human-AI combinations performed worse than the best individual decision-maker on average for decision tasks 34 .”
- Also stated in: § Phase 3: Integrated Practice (Collaboration Mode) · plain
“Vaccaro et al .'s meta-analysis of 106 experimental studies found that human-AI combinations frequently underperformed the best individual decision-maker on clinical decision tasks 34 .”
What the sources say
- Supportedcites [34] Vaccaro, M., Almaatouq, A. & Malone, T. When combinations of humans and AI are useful: a systematic review and meta-analysis. Nat. Hum. Behav. 8, 2293–2303 (2024). source
The Abstract states that the meta-analysis covered 106 experimental studies. The Results (Heterogeneity of human–AI synergy) report a significantly negative pooled effect for decision tasks, and the Results (Overall levels of human–AI synergy) define the comparison as the better-performing of humans alone or AI alone. Our copy provides no page numbers for these passages. (LLM, conf=0.94)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § Phase 3: Integrated Practice (Collaboration Mode) · plain
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“These attributes may develop through sustained engagement with difficult problems in which easy answers are unavailable 37 .”
What the sources say
- Partly supportedcites [37] Ainge, L. E., Edgar, A. K., Kirkman, J. M. & Armitage, J. A. Developing clinical reasoning along the cognitive continuum: a mixed methods evaluation of a novel Clinical Diagnosis Assessment. BMC Med. Educ. 25, 31 (2025). sourceExperimental — accuracy not yet measured
The Discussion section, “Patient perspective and contextual cues simulate ambiguity and diagnostic uncertainty” (page unavailable), says that challenging students to manage diagnostic uncertainty prepares them for complex workplace scenarios. The Background also proposes that such tasks can improve awareness of reasoning strategies. The paper does not establish the claim’s specific mechanism of development through sustained engagement over time. (LLM, conf=0.78)
hedgedstates a possible cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
71 more claims
Create a free account to see all 178 claims and to verify the 3 references still open.
- It may state two separate claims as one. 30
- Probably not a substantial claim. 88
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Competing Interests:
“The authors declare no competing interests.”
§ Disclosure :
“Dr. Pearse A. Keane is a cofounder of Cascader Ltd. and has acted as a consultant for Skleo Health, insitro, Retina Consultants of America, Roche, Boehringer-Ingleheim, and Bitfount and is an equity owner in Big Picture Medical.”
Citations
Cites
Cognitive Load During Problem Solving: Effects on Learning
Sweller, J. (1988) · Cognitive Science · doi:10.1016/0364-0213(88)90023-7 · Cited by 11,048 · cited 2×
“Schema formation requires effortful processing 23 .”
The role of deliberate practice in the acquisition of expert performance
Ericsson, K. A., Krampe, R. T., & Tesch-Römer, C. (1993) · Psychological Review · doi:10.1037/0033-295x.100.3.363 · Cited by 9,564 · cited 1×
“Deliberate practice theory identifies effortful, feedback-rich problem-solving as the mechanism through which expert reasoning develops 22 .”
Rajpurkar, P., Chen, E., Banerjee, O., & Topol, E. J. (2022) · Nature Medicine · doi:10.1038/s41591-021-01614-0 · Cited by 3,212
Artificial intelligence in healthcare: transforming the practice of medicine
Bajwa, J., Munir, U., Nori, A., & Williams, B. (2021) · Future Healthcare Journal · doi:10.7861/fhj.2021-0095 · Cited by 1,856
AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking
Gerlich, M. (2025) · Societies · doi:10.3390/soc15010006 · Cited by 1,248
AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking
Gerlich, M. (2025) · doi:10.2139/ssrn.5082524 · Cited by 1,238
Automation bias: a systematic review of frequency, effect mediators, and mitigators
Goddard, K., Roudsari, A., & Wyatt, J. C. (2012) · Journal of the American Medical Informatics Association · doi:10.1136/amiajnl-2011-000089 · Cited by 1,162 · cited 1×
“The result is a cognitive moral hazard: epistemic responsibility is abdicated not because delegation is appropriate, but because the trainee cannot distinguish appropriate delegation from uncritical acceptance 29 .”
The ethics of AI in health care: A mapping review
Morley, J., Machado, C. C. V., Burr, C., Cowls, J., Joshi, I., Taddeo, M., & Floridi, L. (2020) · Social Science & Medicine · doi:10.1016/j.socscimed.2020.113172 · Cited by 1,105 · cited 1×
“Rather than developing as autonomous reasoners, trainees may come to view themselves as intermediaries who interpret and relay AI outputs 38 .”
Fletcher, G., Flin, R., McGeorge, P., Glavin, R., Maran, N., & Patey, R. (2003) · British Journal of Anaesthesia · doi:10.1093/bja/aeg112 · Cited by 958 · cited 1×
“In medical education, nonspecialists who stand to benefit most from AI support are simultaneously those most susceptible to overreliance 32,33 .”
When combinations of humans and AI are useful: A systematic review and meta-analysis
Vaccaro, M., Almaatouq, A., & Malone, T. (2024) · Nature Human Behaviour · doi:10.1038/s41562-024-02024-1 · Cited by 570 · cited 1×
“Evidence on collaborative performance is variable 34,35 , and AI augmentation can meaningfully extend clinical capability in well-defined tasks.”
Research works printed but not linked: 2
- American Medical Association. 2 in 3 physicians are using health AI—up 78% from 2023. https://www.ama-assn.org/practice-management/digital-health/2-3-physicians-are-using-health-ai-78-2023 (2025).
- Ten Cate, O. Competency-based postgraduate medical education: past, present and future. GMS J. Med. Educ. 34, Doc69 (2017).
Cited by
AI Assistance Reduces Persistence and Hurts Independent Performance
Liu, G., Christian, B., Dumbalska, T., Bakker, M. A., & Dubey, R. (2026) · arXiv · doi:10.48550/arxiv.2604.04721 · Cited by 3
“…persistence and independent ability for all tasks; nonetheless, at minimum, our findings suggest that AI assistance displaces the very practice opportunities through which people develop lasting competence, a concern echoed in recent perspectives on AI use in medical training (Ke et al., 2026).”
Grunebaum, A. N., Dudenhausen, J., & Chervenak, F. (2026) · American Journal of Obstetrics and Gynecology · doi:10.1016/j.ajog.2026.06.011 · Cited by 4
Nam, Y., Jeong, J., & Kim, N. (2026) · arXiv · doi:10.48550/arxiv.2606.09500 · Cited by 4
Building AI companions that prioritise learning over performance
Khosravi, H., Gašević, D., Sadiq, S. W., Yan, L., Lodge, J. M., Tangen, J. M., Denny, P. C., DiCerbo, K. E., Shum, S. B., & Baker, R. S. (2026) · Computers and Education Artificial Intelligence · doi:10.1016/j.caeai.2026.100680 · Cited by 9
From Flexner to artificial intelligence: 100 years of Academic Medicine
Byington, C. L. (2026) · Academic Medicine · doi:10.1093/acamed/wvag178
Metadata
- Authors
- Ke, Y., Jin, L., Ong, J. C. L., Thirunavukarasu, A. J., Car, J., Cheung, C. Y., Tham, Y. C., Ting, D. S. W., Ong, M. E. H., Compton, S., Narayan, A., Keane, P. A., Wong, T. Y., Bates, D. W., Tan, P., & Liu, N.
- Affiliations
- Singapore General Hospital (SG)
- Singapore National Eye Center (SG)
- London School of Hygiene & Tropical Medicine (GB)
- King's College London (GB)
- Chinese University of Hong Kong (HK)
- National University of Singapore (SG)
- SingHealth (SG)
- Duke-NUS Medical School (SG)
- Duke University (US)
- Moorfields Eye Hospital NHS Foundation Trust (GB)
- Brigham and Women's Hospital (US)
- Agency for Science, Technology and Research (SG)
- Corresponding author
- Liu, Nan
- ORCID
- Venue
- Nature Medicine
- Published
- 2026-05-22
- Volume
- 32
- Issue
- 6
- Pages
- 1997-2006
- ISSN
- 1078-8956 (ISSN-L)
- 1546-170X
- OpenAlex
- W7162142312
- Semantic Scholar
- 42884cd2cacd804a17fde1ee7d0374c4fae674e6
- Cited by
- 71
- References
- 38
- Citations by year
- 2026: 71
- FWCI
- 32.93
- Type
- journal article
- Language
- en
- Open access
- green
- doi.org
- Licence
- www.springernature.com/gp/researchers/text-and-data-mining (tdm)
- www.springernature.com/gp/researchers/text-and-data-mining (vor)
- Locations
- Nature Medicine (publishedVersion)
- SSRN Electronic Journal (submittedVersion)
- PubMed (publishedVersion)
- LSHTM Research Online (London School of Hygiene and Tropical Medicine) (acceptedVersion)
- Research Portal (King's College London) (submittedVersion)
- Funders
- Duke-NUS Medical School
- Topics
- Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
- Social Sciences › Education › Innovations in Medical Education
- Health Sciences › Medicine › Family Practice › Clinical Reasoning and Diagnostic Skills
- Keywords
- artificial intelligence in medical education, clinical reasoning, medical students, deskilling, AI-generated content, competency assessment, metacognition, professional identity formation, health care equity, medical licensing
- MeSH terms
- Artificial Intelligence, Clinical Competence, Education, Medical, Humans, Learning
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: AI-induced never-skilling in medical education. https://app.cochat.ai/papers/10.1038/s41591-026-04438-y
- The paper
- Ke, Y., Jin, L., Ong, J. C. L., Thirunavukarasu, A. J., Car, J., Cheung, C. Y., Tham, Y. C., Ting, D. S. W., Ong, M. E. H., Compton, S., Narayan, A., Keane, P. A., Wong, T. Y., Bates, D. W., Tan, P., & Liu, N. (2026). AI-induced never-skilling in medical education. Nature Medicine. https://doi.org/10.1038/s41591-026-04438-y
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.