LLMs for medical education
The utilisation of AI systems extends beyond ethics, and many studies have highlighted their applications in diverse domains of education [7, 23, 24]. Selin Akgun and Christine Greenhow [24] outline four key benefits of using AI in education: (1) personalised learning systems; (2) automated assessment; (3) facial recognition with predictive analytics; and (4) social networking services, including chatbots. As a form of LLMs, chatbots have been recognised as an effective tool for facilitating active learning in students. Research in this domain, such as studies evaluating the efficiency of various chatbots in language learning [25], demonstrates the potential of LLMs in enhancing educational experiences through personalised learning and reduced dependency on traditional resources.
In medical ethics education, significant research predates the introduction of LLMs. In the 2000s, Michael Anderson and colleagues developed MedEthEx, an expert system consolidating ethical judgements from healthcare professionals to provide general ethical guidelines [26]. This system was pivotal in advancing medical ethics education. Additionally, proposing the concept of an artificial moral advisor (AMA) for moral enhancement, Alberto Giubilini and Julian Savulescu [27] argued that AI, as a consultative partner in ethical decision-making, could enhance the quality of ethical judgements, and that an AMA ‘could facilitate education, growth, and moral development’ by presenting diverse ethical perspectives (p. 175). While these discussions primarily focused on AI-assisted moral judgement in general and do not explicitly recommend using LLMs as instructional tools, they offer valuable insights into the potential role of LLMs in virtue cultivation within medical ethics education. Two questions are pertinent here: what role can LLMs play in virtue cultivation, and how should students engage with LLMs for this purpose?
LLMs as exemplars
A key focus of comprehensive ethics education is cultivating virtue. To use LLMs for virtue cultivation, defining the concept of virtue is crucial. We draw upon the theory of exemplarism, as articulated by Linda Zagzebski [28]. Exemplarism posits that understanding virtues, right actions, and duties requires direct reference to exemplary individuals. In this framework, ethical exemplars are not merely good individuals but paragons of virtue. Their actions serve as a model for ideal behaviour. To cultivate virtues, learners need to recognise and emulate these exemplars.
This approach to virtue cultivation is profoundly rooted in Aristotle’s Nicomachean Ethics [29]. Aristotle proposes that virtue is defined by the discernment of a prudent individual, highlighting exemplary persons as moral standards. Although he views habit, rather than imitation, as central to virtue formation, Zagzebski extends Aristotle’s concept of mimesis − originally outlined in Poetics − to encompass not just creativity but all virtues. As she writes:
Mimesis, or imitation, precipitates virtuous behavior, and the basic human pleasure in mimesis provides the initial motivation to act as virtue requires. But mimesis leads to the performance of a virtuous act for its own sake because what is aimed at in the mimesis of an activity is just the activity [28]. (p. 134)
While various theories of virtue acknowledge the importance of models or exemplars in teaching virtues, exemplarism specifically emphasises their use, making it particularly relevant for integrating LLMs into virtue education. An effective LLM for this purpose should not only discuss ethical dilemmas but also model virtuous behaviours in those discussions. Drawing from neo-Aristotelian philosophy [30], the concept of ‘v-rules’, such as ‘act compassionately’ or ‘act honestly’, can guide the design of these LLMs. These guidelines enable LLMs to demonstrate actions characteristic of virtuous individuals, thereby going beyond standard responses to ethical dilemmas and illustrating virtuous behaviour in various scenarios (Fig. 1). In fact, some recent attempts to develop LLMs have focused on their potential to facilitate the cultivation of virtues. Charlene Tan proposed, for example, the concept of Digital Confucius, an AI agent modelled on the teachings of Confucius, as a means to impart virtue.
Ethical exemplar in medicine. This diagram illustrates the core elements of the ‘Ethical Exemplar in Medicine’ framework, which aims to enhance the development of essential virtues for practicing medicine. The central icon represents the concept of moral exemplarism, surrounded by guiding questions that foster critical reflection and dialogue on virtuous behaviour in medical practice. These questions address practical scenarios, diverse perspectives, and the nuances of compassion and ethical decision-making, encouraging healthcare professionals to align their actions with the principles of virtue ethics
LLMs as advisors
For LLMs to serve effectively as tools for teaching virtue, students must appropriately engage with them. Treating the ‘virtuous’ attitudes generated by LLMs as unquestionable testimony or evidence (e.g. ‘We should follow the LLM’s answer!’) would signify a failure in ethical education. Alison Hills [32] points out that there is a crucial distinction between considering others’ claims as advisory and accepting them as absolute evidence or testimony: to put it from our perspective, we can regard the exemplar as an advisor or as a witness. Ethical education should encourage students to regard LLMs as an advisor and critically engage with and reflect upon the answers from LLMs, thereby examining more deeply what is virtuous in the related contexts rather than simply shifting responsibility onto the LLMs.
In LLM-based ethical education, students must recognise that LLMs serve as advisors of virtues; they are not definitive authorities. LLMs can effectively teach virtues by guiding students through detailed scenarios and suggesting specific behavioural responses. Students should focus on understanding the process by which LLMs arrive at certain decisions, analysing the details of situations and the perspectives considered. This encourages students to discern which aspects of the LLM’s guidance are valuable for emulation and which are not, fostering critical engagement and independent ethical thinking.
Achieving the objectives of using LLMs in ethical education hinges significantly on the nature of prompts given to the LLM. Vasiliki Rahimzadeh and colleagues [10] observed that when they input the prompt ‘complete an ethical workup’ for a complex medical scenario into ChatGPT, the response was primarily grounded in Beauchamp and Childress’ four principles (p. 23). Responses based on alternative frameworks, such as virtue or feminist ethics, emerged only with explicit prompting. Rahimzadeh and colleagues noted: ‘ChatGPT’s response implies a hierarchy of ethical frameworks atop which principlism is presumed to sit’ [10] (p. 23). This suggests that the design of prompts greatly influences the LLM’s output, especially in the context of virtue cultivation.
In fact, even current versions of ChatGPT can generate responses focusing on empathy and moral sensitivity (Fig. 2). However, as Rahimzader and colleagues pointed out, it is only realised by a rather careful selection of prompts. Therefore, when students in need of ethical guidance enter simple prompts into current LLMs, which are not tuned for medical ethics education, there is a substantial risk of unintentionally perpetuating unsuitable behavioural paradigms. Hence, we do not believe that LLMs such as ChatGPT, in their present forms, could be considered an exemplary model or advisor for ethical education.
Compassionate approaches in breaking bad news: The Case of Mr. A. This figure provides a detailed example of Chat GPT’s recommendations of compassionate communication in the doctor-patient relationship, specifically focusing on Mr. A, a patient diagnosed with stage 4 gastric cancer. The strategies provided are divided into two core components. First, ‘Breaking Bad News with Sensitivity’ is explained interms of empathy, and clarity, and honesty. For empathy, ChatGPT provides a specific example of supportive and understanding language: ‘This is difficult news to share, and I want you to know I’m here to support you through this’. Second, ‘Acknowledging Emotional Reactions’ is presented as validating and supporting the patient’s emotional response. Validation is explained with specific phrases ‘It’s natural to feel overwhelmed right now’. To support the patient’s response, the importance of patience is also acknowledged to give the patient time to absorb the news. This example highlights that with appropriate settings, it is possible for the current LLMs to give recommendations in terms of virtues
To fulfil the objectives outlined previously, it is crucial to rigorously refine LLMs, prompts, and human critical engagements. We believe, however, that it is safe to be optimistic about the possibility of these improvements. For example, the model depicted in Fig. 2, despite being subjected to only basic fine-tuning, manages to provide advice that might plausibly act as a preliminary framework for ethical behaviour. It is expected that subsequent, more comprehensive experiments conducted by experts will markedly improve the capability to produce superior educational materials.
As argued above, even with regards to fine-tuned LLMs, educators should encourage students to critically examine the answers generated by LLMs. While this emphasises the importance of human resources in cultivating virtue, however, what is important for our arguments is that it is clearly less burdensome for teachers to support students’ use of LLMs than for teachers to carry out all the related processes themselves.
Thus, incorporating LLMs into the medical ethics education framework constitutes a promising approach to enhance the pedagogical landscape. This method offers a means to integrate non-principlist elements into the educational environment without making excessive demands on human and temporal resources. Consequently, LLMs can serve as a crucial tool in ethical education, facilitating a level of depth and sophistication in the learning experience previously considered prohibitive due to resource constraints.
