This Week in MedEd
16 open funding calls, 57 upcoming events, 60 live jobs (21 new, 11 closing this week), including:
💰 Digital Skills and Jobs Platform: National Coalitions for Digital Skills and Jobs (up to €2M from the European Commission’s Digital Europe Programme, deadline 1 Oct 2026)
🎓 2027 AAMC Medical Education Innovation Conference (Indianapolis, IN, 17-20 Apr 2027)
💼 Assistant/Associate Professor in AI for Medical Education (Duke-NUS Medical School, Singapore)
The full list, with deadlines, links and eligibility, is one tap away. Try everything free for 30 days.
Key points
Across four randomised trials, AI-assisted surgical skills training was statistically indistinguishable from expert human instruction.
The apparent benefit of AI training shows up only when the comparator is doing very little.
96.3% of medical and health science students at one UK school had used generative AI academically, but only 40.0% had received any formal training in it.
Students who had been trained reported significantly lower trust in AI outputs, which is probably the point of training.
A 30-author international working group has published seven domains and 24 learning objectives, and is asking the ACGME and national Royal Colleges to make AI competency a required and assessable element of training.
Cai et al. undertook a systematic review and meta-analysis of studies that looked at AI-assisted training for operative technical skills1. The review surfaced nine studies comprising 588 participants. The headline pooled effect was not significant and wide, but the subgroup analysis is very interest interesting:
Against active human instruction the pooled effect was 0.13 (95% CI −0.24 to 0.51, P = 0.48).
Against passive conditions or training without AI it was 0.94 (95% CI 0.30 to 1.59, P = 0.004).
So it seems that the apparent advantage of AI-assisted training is greatest when the alternative is not being taught very much. Set against a competent human teacher, it disappears.
There is a caveat, though. The certainty of evidence was rated low under GRADE, downgraded for risk of bias and imprecision, and to their credit the authors, they call the passive-comparator result "exploratory" rather than presenting it as a settled finding.
Effiong and Effiong surveyed 128 medical and health science students at the University of Birmingham2. They found that 96.3% of students were using genAI, but only 40.0% reported receiving any formal training. This finding is not particularly illuminating, and confirms what we already know from much larger national survey, such as the HEPI Student Generative Artificial Intelligence Survey.
There was another finding, which has also been reported previously, that caught my eye: students who had received training reported significantly lower trust in AI outputs (Mann-Whitney U = 321.50, p = .004, surviving Bonferroni correction).
This is super interesting to me. It suggests that training does not produce enthusiasts, it produces sceptics. I think we are beginning to build up a picture of factors that are protective against AI slop. In addition to additive training, there might also be personality traits that contribute to this. I previously wrote about this concept during the spring:
Rosella et al. ask "what should we actually teach" in relation to GenAI3. Writing in npj Digital Medicine, an International AI in Medicine Education Working Group drawn from 23 universities and healthcare centres sets out seven domains and 24 learning objectives, aimed at undergraduate medical education but adaptable to other contexts. The domains are:
fundamentals
quality of evidence
responsible healthcare AI
deployment, post-deployment monitoring, evaluation and oversight
leadership and change management
anticipating future trends
case studies.
The first three build appraisal skills, the next three extend into implementation, and the last integrates everything through case-based reasoning.
The report is refreshingly honest. They report real disagreement about how much technical depth clinicians need, settling on conceptual fluency because "the aim is not to train all clinicians to become AI developers". They also found responsible AI was "the domain where the widest variation in curricula existed", with some programmes running standalone ethics modules and others barely addressing it.
They close by asking the ACGME and national Royal Colleges to treat AI competency as required and assessable. As with many of these ‘frameworks’ papers, its worth noting that this is expert consensus rather than a formal Delphi, and their recommendations have not been validated in any way.
I’ve written before about my frustration at the absolute glut of data-lite frameworks in AI x MedEd literature and conferences. Ghavami Hoseinpour et al. have produced another one by synthesising 25 articles into a five-domain ethical framework4:
ethical foundations
educational integration and curriculum governance
clinical training and human oversight
assessment and academic integrity
data governance
While the framework is sensible, I'm not sure it adds much. The useful part of this paper, and the reason I've included it here, is their critique of frameworks in general. They find that "implementation strategies remain largely abstract, with minimal empirical validation", that institutions risk "ethics washing" by endorsing principles without building accountability, and that the field keeps framing algorithmic bias as a technical flaw rather than "a reflection of broader institutional, historical, and sociocultural inequities embedded within medical education systems themselves".
They also note that student perspectives are nearly absent from this literature, which they attribute to learners being positioned as passive recipients of systems designed for them.
Ironically the authors also say that their own framework "remains a preliminary conceptual model that has not yet undergone formal validation", which is bold for a paper criticising unvalidated frameworks!
Based on these papers I think the problem is reasonably clear. Cai et al. show that the best randomised evidence we have cannot distinguish AI instruction from a good teacher. Effiong and Effiong (and lots of other studies before them) show that near-universal adoption happened anyway, with the majority untrained. Rosella et al. and Ghavami Hoseinpour et al. argue that AI use should be governed as an ongoing commitment rather than a one-off.
If AI's demonstrated advantage is over poor teaching rather than good teaching, then every deployment decision is really a question about what it replaces. Swapping a supervised session for a chatbot is not supported by this evidence. Filling the hours between timetabled teaching, when most students currently get nothing at all, probably is.
Cai M, Luo B, Xiang Q, Wang P. Artificial intelligence-assisted surgical skill acquisition: a systematic review and meta-analysis of randomized controlled trials. BMC Medical Education. 2026. https://doi.org/10.1186/s12909-026-10257-z
Effiong I, Effiong CJ. Generative AI, social normalisation, and the governance gap: a cross-disciplinary study of AI adoption and sustainability in UK medical education. BMC Medical Education. 2026. https://doi.org/10.1186/s12909-026-10330-7
Rosella LC, Aphinyanaphongs Y, Barry J, et al. A global framework for artificial intelligence education in medicine: international working group recommendations. npj Digital Medicine. 2026. https://doi.org/10.1038/s41746-026-03197-x
Ghavami Hoseinpour B, Karimian Z, Zarifsanaiey N, Hooshmandja M, Emadi N. Toward a context-specific ethical framework for artificial intelligence in medical education: a critical narrative review informed by a systematic literature search. BMC Medical Education. 2026. https://doi.org/10.1186/s12909-026-10240-8