‘Slowing down when you should’: optimising the translation of artificial intelligence into medical school curricula

INTRODUCTION This is an exciting and anxious time for medicine and medical education as innovations and applications of artificial intelligence (AI) in both domains proliferate at a rapid, dizzying pace. In this article, we call for a considered approach to the design and implementation of AI usage by students in undergraduate medical education (UGME). To do so, we adopt the metaphor of 'slowing down when you should' from the intraoperative surgical decision-making literature,1 where surgeons mobilise cognitive resources during moments in surgery to avoid errors and choose the best course of action. Like surgery, this would mean that 'moments' should be taken during the implementation of AI into medical school curricula to assess and plan subsequent steps. We briefly explain in this short article why doing so is useful. Second, and following our first point, we propose adopting frameworks from Implementation Science (IS) to guide the successful incorporation of AI teaching into medical education, as this enables consideration of best evidence medical education practices while being mindful of patient safety concerns. THE CONTEXT: ARTIFICIAL INTELLIGENCE, MORAL PANICS AND THE TECHNICAL FIX The speed of movement in the adoption of any given technological innovation is often suffused by the seduction of originality, progression and the promise of mastery over current and future problems. AI is no exception — although it is exceptional in terms of its ubiquity and pace of adoption — to the point that we, in medical education, are in danger of taking on the approach of a 'solution in search of a problem', rather than identifying the problem and then finding a fit-for-context solution.2 is is known as the 'technical fix'3 — a belief that ongoing pernicious educational problems can be instantly solved through the application of technology without considering possible unintended consequences. The teaching of AI in medical education fits squarely into this type of conundrum. For instance, there has been much discussion in both popular media and academic forums about the hypothetical and actual problems associated with AI adoption and implementation in education and clinical practice. For example, there are concerns that large language models, such as ChatGPT, are creating artificial scholarship in academia4 and media stories about unethical behaviour of students and institutional responses (some of which are problematic).5,6,7,8,9 In clinical practice, there are concerns over the creation of hallucinations (false information) that undermine diagnostic accuracy, threats to privacy and the perpetuation of bias and racism.10 We face an additional complicating factor in medical education. Studies of large language models in medical consultations are purportedly revealing weaknesses in physician competency. A recent article published in JAMA compared AI chatbot and physician responses to patient questions posted on a social media forum. The chatbot responses were rated higher for both quality and empathy. However, it only takes a quick scan of the comments associated with this study to see that it contains potentially serious methodological flaws.11 Unfortunately, such misleading articles add weight to the construction of a moral panic — 'a condition, episode, person or group of persons emerges to become defined as a threat to societal values and interests12' — concerning AI in medicine and medical education. This threat is often exaggerated.13 The crucial point we want to emphasise is that there are plenty of real and imagined pressures creating a rush to implement a curriculum that involves the adoption of AI by students, and/or teaching about the use of AI in health care within medical schools to prepare workplace and future-ready physicians. This anxiety that is leading to a rush to implement AI may also lead to erroneous decisions about when, how and what kind of AI curriculum we should be implementing in medical education. Like our colleagues in surgery, medical educators need to exercise some situational awareness of the new emerging context in which we are operating and 'slow down when we should', to best develop and implement AI teaching and learning. To help in the quest for a measured approach to AI curriculum development, we propose an agenda informed by IS, as outlined in the following section. What we know: current state of AI teaching in medical education In 2021, two published reviews identified gaps and key themes in the peer-reviewed literature on AI training in medical schools. Lee et al.'s14 scoping review reinforces the need for AI curriculum and makes key recommendations for curriculum content (i.e., machine literacy and ethics) and delivery (i.e., a focus on empathy training to combat the potential dehumanising effects of AI usage) and, finally, the barriers and facilitators of the implementation of AI curriculum in medical schools. The studies reviewed exhibited a high degree of heterogeneity and poor consensus concerning curricular content and delivery. Nonetheless, five areas of curricular delivery were identified: (a) working with and managing AI systems, (b) ethical and legal implications of AI systems, (c) a continued emphasis on biomedical knowledge and pathophysiology of disease, (d) critical appraisal of AI systems, and (e) working with electronic health records. The second review in 2021 by Grunhut et al.15 also found a paucity of plans and/or reports on the implementation of AI teaching in medical schools, and similarly recommended the need for more research in this area. While giving useful guidance, neither review provides solutions on how to overcome the significant barriers that remain when introducing any new teaching, such as time and space in a packed curriculum, faculty development and logistical support.14,15 We, therefore, suggest using IS interventions to help us 'slow down when we should' in respect to identifying areas where the implementation of AI in medical education might be most useful. What to do once the evidence is built: Implementation Science in medical education The thinking and practice of IS (also known as knowledge translation) is no stranger to medical education. Three significant articles outline what IS is,16 the need for the use of IS within medical education thinking and practice,17 and the building of an implementation research agenda in medical education.18 Although this work pre-dates the advent of AI implementation in medical education, the IS approach is apt and instructive in terms of managing the design and delivery of AI teaching. Implementation science involves the dissemination of innovative solutions throughout a system, once a needs assessment has been completed and evidence base established. Many theories and methodologies constitute IS.17 Pertinent to this article, Carney et al.18 translated Damschroder et al.'s19 consolidated framework for implementation research (CFIR) to fit the medical education environment. At heart, CFIR determines how key elements must be identified and modified to help an innovation move from adoption failure to success Figure 1.Figure 1: Diagram displays an adapted conceptual framework for implementation research in medical education. Taken from Carney et al.[ 18 The article is distributed under the terms of the Creative Commons Attribution 4.0 licence, which allows others to copy and redistribute the material in any medium or format and to remix, transform, and build upon the material for any purpose, even commercially, as long as appropriate credit is given (https://creativecommons.org/licenses/by/4.0/deed.en)].There are three major challenges to overcome as we implement AI teaching: (a) a need for significant investment in faculty development in the study and application of IS and AI methods in clinical education settings; (b) funding to support these activities for both universities and clinical care settings, and more broadly; and (c) techniques used in IS need to be underpinned by appropriate theoretical frameworks to ensure that the social aspects of change are identified and evaluated.18 This last point is directly relevant to acknowledging the context of AI adoption as partially shaped by a moral panic around AI, the effects of which are compounded by the allure of the 'technical fix' and the notion of 'solutions in search of problems'. Being aware of these issues means we need not be driven by them. Instead, we propose that adopting IS is crucial to developing competencies that allow for deliberately curated (inter)actions with AI by students and teachers. This approach enables 'slowing down' to properly assess our medical education AI needs, the broader context and the local environment. Understanding these contextual factors is an essential first step in shaping AI teaching implementation to give it the best chance of being accepted and embedded. This raises the following question: How do we manage this process in a context where the evidence of AI teaching and learning will continue to evolve rapidly and medical schools will face socio-political pressure to respond quickly, oftentimes without evidence? A halt in the adoption of AI in UGME is not a viable option. We suggest that the design and implementation of AI teaching can be informed by the adoption of an ongoing 'backwards planning' logic model that enables continual progression in the absence of any guidance from a standardised competency-based curriculum. 'Backwards planning': building on real-world clinical scenarios in medical schools In 2023, Krive et al.20 took on the challenge put forward by Lee et al.14 for further research to create a framework of competencies to underpin the adoption and implementation of a standardised AI curriculum. They adopted a backwards design principle in the design of online learning assignments simulating real work done in the clinical environment. Their aim was to get fourth-year medical students ready for the new digital patient care environment. They outlined a modular 4-week AI course for the students that integrated AI with evidence-based medicine, pathology, pharmacology, tele-monitoring, quality improvement, value-based care and patient safety.20 This resulted in demonstrable AI knowledge and skill acquisition among the students. The authors also outlined how their course could be integrated into medical education. The significance of this work lies in the foundations of the design. Krive et al. worked with simulated scenarios drawn from real-world AI usage in the clinical environment, that is, scenarios that were steeped in immediate clinical relevance. This kind of 'backwards design' curriculum construction principle20 can bring in the complexity of working in the clinical environment to help support medical students as they transition into clinical practice. In essence, it may enable a practical engagement with the teaching of AI to medical students while the evidence base for a standardised curriculum is being developed. However, there is a caveat, or unintended outcome of teaching AI: the potential emergence of an uncritical over-reliance on AI by students as they enter the clinical training environment. Artificial intelligence and training for (over)certainty: an immediate priority We know that medical students want to achieve control over their clinical learning and practice. The issue is that, in their rush to do so, they may neglect the complexity of clinical decision-making and reduce their ability to evolve their learning strategies. How then does this relate to adoption of AI in medical education by learners? Over the last 60 years, research on medical uncertainty has provided three key insights into what unintended effects might occur once medical students begin interacting with AI in the clinical environment: (1) the impossibility of mastery over constantly evolving medical knowledge and skills; (2) uncertainty about gaps in the learner's medical knowledge; and (3) judging whether the problems they encounter in the clinic are the result of an individual skill or knowledge deficit, or a failure of medicine itself.21,22,23 In short, as medical students enter the clinical environment, unless they are taught appropriately, many are likely to uncritically embrace AI as another tactic to gain control over medical uncertainty. Protecting against this requires a shift in focus from the speed of adoption to appropriate adoption of AI. Clinical educators should focus on developing student/trainee competencies that allow for a critical engagement with AI during crucial transition periods in medical training (i.e. the transition from campus to clinic, from medical school to postgraduate training). Drawing on Light's work,23 this means teaching medical students how to foresee and account for the possibility of errors in the form of hallucinations and other shortcomings of nascent AI tools.10 This places an even greater sense of urgency on the previously identified need for the design and implementation of AI critical literacy as a key core competency within medical school curricula.14,15,20 CONCLUSION In this article, we have proposed an approach to 'slowing down when we should' in AI teaching design and implementation. 'Slowing down' and adopting IS thinking and practice open up new ways of thinking that can optimise the process and outcomes of bringing AI into medical school curricula. Financial support and sponsorship Nil. Conflicts of interest There are no conflicts of interest.