What we know about what trainees know
Five 2026 studies on AI literacy, adoption, placement use, and the cost of dependence.
Hello everyone!
Many of you will remember that academics used to have something called ‘summer’, which was a break between academic years when we could planning, think and write about things that we didn’t have time to do during periods of busy teaching. Most of us don’t have much of a summer any more. In my University, assessments run later into June, and clinical students resume their studies in August or even July.
If you have managed to do some writing this year, I thought you might find it useful to have access to my list of target journals so you can identify where might be a good place to publish. This includes dozens of journals, their impact factor and their publication model (i.e. open access, hybrid etc).
If 2024 was the year we started worrying about whether students were ready for AI, 2026 is shaping up as the year we actually find out. The vibes-based debate is giving way to validated scales, scoping reviews, configuration analyses, and survey data from actual placements. We finally have something that looks like a map.
This week I’m covering five recent papers that sketch a fuller picture of where medical and health professions students actually sit with AI. They cover what we think they should know, how we might measure what they do know, what shapes their willingness to adopt these tools, what they are already doing during clinical placements, and what overuse may be costing them.
Key points:
A new scoping review of 54 studies from 22 countries finally puts a structure on the dozens of fragmented “AI competencies for medical students” lists circulating since 2022.
We now have a validated, theory-grounded AI literacy scale for medical students. Until now, most assessments were generic or borrowed from other disciplines.
Two conditions appear necessary for medical students to form strong intentions to adopt AI: belief in its future clinical relevance, and a positive attitude toward its inclusion in the curriculum. Without both, technical familiarity alone does not move the needle.
More than half of health professional students at one French university used generative AI during clinical placements. Adoption varies by discipline and tracks closely with self-perceived readiness, not formal training.
AI dependence is correlated with measurable increases in anxiety and depression in medical students. Each 1-point increase on a dependence scale is associated with roughly a 5% rise in expected anxiety scores.
Hunt et al. offer the first attempt I’ve seen to bring order to the chaotic landscape of AI competencies for medical students1. Their PRISMA-ScR scoping review of 4071 records yielded 54 included studies from 22 countries, from which they distilled 564 verbatim competency statements into a structured taxonomy of seven domains, 37 competencies, and 170 learning objectives.
The seven domains span AI ethics, AI law and regulation, AI professionalism, clinical applications, critical appraisal of AI output, research and innovation, and the theory and foundations of AI. The authors are honest about a major weakness in the underlying literature: most source documents are editorial or opinion pieces rather than empirical curriculum studies, and emphasis clusters heavily on ethicolegal oversight and critical appraisal. As a curriculum planning tool this is a useful starting point, but it is more a structured inventory of what people think doctors should know than evidence about what they actually need.
Lin et al. address a closely related problem: even if we agree on what AI literacy should look like, how do we measure it2? Their study develops and validates the AI Literacy Scale for Medical Students (ALSMS) within a self-determination theory framework, using a split-sample design with 518 medical students (n=204 for exploratory factor analysis, n=314 for confirmatory factor analysis). The final structure has nine factors organised into the SDT domains of competence, relatedness, and autonomy, with strong psychometric properties.
The choice of SDT is itself an editorial statement: it positions ethical and autonomous use as core to what literacy is, not as a soft add-on. For programmes evaluating their AI curricula, this is the first instrument I’ve seen that is purpose-built for medical students and grounded in a coherent learning theory.
Ringeval et al. then ask the obvious follow-up: what conditions actually drive students to want to use AI in their future practice3? Using Necessary Condition Analysis and fuzzy-set Qualitative Comparative Analysis on survey data from 177 students at a Canadian medical school, they identify two conditions that are necessary (not merely helpful) for strong adoption intentions: a belief in the future relevance of AI-based health technologies to medical tasks, and a positive attitude toward including these technologies in the curriculum.
The fsQCA further surfaces two distinct configurations sufficient to foster adoption, which the authors describe as “AI profiles.” For curriculum leads, the practical message is that hands-on familiarity alone is insufficient. If students do not believe AI will matter in their future practice, or do not see it as legitimately part of medicine, technical exposure will not carry them across the line.
Kotzki et al. ground the conversation in what is already happening on the wards. Their cross-sectional survey of 388 health professional students at a French university (mostly nursing, but spanning medicine, pharmacy, midwifery, and physiotherapy) found that 52.6% had used generative AI during clinical placements4.
Use varied significantly by discipline and rose sharply with self-perceived GenAI maturity, from 9% in the lowest-maturity group to 76% in the highest. Among users, the dominant tasks were information retrieval (77.9%), bibliographic search (74.5%), and translation or rephrasing (71.1%); patient-facing uses such as document drafting and communication preparation were much less common. More uncomfortably, 23.5% of users reported entering patient-identifying information at least once, and 47.1% had processed real medical content they perceived as anonymised.
Students themselves flagged dependency (90.9%), skill erosion (84.8%), and confidentiality breaches (87.4%) as the top risks, and wanted ethics or regulatory training (77.7%) and a clear best-practice clinical guide (78.3%) above almost everything else. The honest reading is that GenAI is already informally embedded in placement practice without supervision or guidance, and the absence of governance is generating both real privacy risk and a hidden curriculum about what “normal” use looks like.
Chavez Sosa and Huancahuire-Vega complete the picture with a study of the psychological cost of leaning too hard on these tools5. In a cross-sectional sample of 187 medical students at a Peruvian university, they applied a validated AI Dependence Scale alongside the DASS-21 measure of stress, anxiety, and depression.
AI dependence correlated with anxiety (ρ=0.336), depression (ρ=0.316), and stress (ρ=0.277), and in adjusted negative-binomial regression models, each 1-point increase on the AI dependence scale was associated with a 5% increase in expected anxiety scores and a 4% increase in depression scores. The single-site, cross-sectional design means we cannot infer direction of effect, and the effect sizes are modest, but the signal is consistent with a growing concern that high-stress trainees may be leaning on AI in ways that mirror other dependence behaviours rather than functioning as a clean productivity boost.
These papers describe a field that is starting to grow up. We have a structured (if still aspirational) taxonomy of what trainees should know, a validated scale to measure where they are, an emerging causal map of what makes them want to adopt these tools, hard data on what they are already doing in clinical placements, and the first careful look at what dependence may be doing to their mental health. None of this is the final word, and the geographic spread (Canada, Taiwan, France, Peru, plus a multi-country scoping review) is itself a useful reminder that the trainee story differs by context. But the era of “we don’t really know what students are doing” is, mercifully, ending.
Hunt, V.M., Sprehe, L.K., de Lomba, W.C. et al. What the AI era doctor should know: a scoping review of proposed artificial intelligence competencies for medical education.npj Digit. Med. (2026). https://doi.org/10.1038/s41746-026-02761-9
Lin, H. C., Lin, C. S., Chai, C. S., Lin, C., Tsai, P. J., & Liang, J. C. (2026). Measuring AI literacy in medical students: scale development and validation within a self-determination theory framework. Medical Education Online, 31(1). https://doi.org/10.1080/10872981.2026.2675066
Ringeval, M., Raymond, L. & Paré, G. Identifying necessary conditions for medical students’ adoption of AI in the future practice: a survey study in Canada. BMC Med Educ (2026). https://doi.org/10.1186/s12909-026-09432-z
Kotzki S, Massonnet Turner C, Gauthier K, Minoves M, Vuillerme N
Health Professional Students’ Use of Generative Artificial Intelligence During Clinical Placements: Cross-Sectional Online Survey Study
JMIR Med Educ 2026;12:e85243
doi: 10.2196/85243
Chavez Sosa JV, Huancahuire-Vega S
Anxiety and Depression Associated With the Dependent Use of Generative AI in Medical Students: Cross-Sectional Study
JMIR Form Res 2026;10:e82667
doi: 10.2196/82667
Thank you Andrew for a summer kick-starter. I am a medical professional designing and integrating AI into current workflows and job roles and you highlight some important points around adoption, engagement and the need for preparing qualified medical profesionals for a future that will become more and more linked to AI applications and dependencies.