Caution against anthropomorphising the virtues

Arslan et al proposed the BODHI (balanced, open-minded, diagnostic, humble, and inquisitive) framework for clinical AI, applying epistemic constraints to counter the ‘unwarranted confidence’ of AI systems and mitigate patient harm.¹ The authors propose that this can be achieved in part by implementing a Virtue Activation Matrix, allowing users to plot medical complexity compared with the model’s confidence in providing medical decision-making.¹ Arslan et al suggest that this is a method for implementing epistemic virtues in the context of AI-delivered clinical care.¹ We commend this work but argue that claims that clinical AI ‘should embody’¹ virtues such as humility and curiosity risk anthropomorphising AI, obscuring distinctions between human moral agency and computation. This could distort public understanding and ethical frameworks governing AI development.

Arslan et al present a view that the simulation of a virtue is interchangeable with the actual virtue. We disagree. Given our backgrounds in philosophy of medicine and bioethics, we do not subscribe to the idea that placing epistemic constraints on AI equates to genuine virtue. In Aristotelian ethics, for example, virtue is rooted in character cultivated through lived experience. Phronesis (practical wisdom) is the contextual ability to discern moral significance and reflect on what it demands. Modern interpretations of humility have expanded on this virtue. Vallor defines technomoral humility, or humility about novel technologies, as an acknowledgement of the limits of our knowledge and abilities as a species, and that we must reject the notion that new technologies push us towards perfect human control of our environment.² In Vallor’s view, engineering an LLM to avoid harm is an external behavioural constraint, not an expression of character. This is clearest with epistemic humility: a humble agent recognises the limits of its understanding from the inside, through self-awareness and an openness to being wrong. An LLM programmed to hedge merely simulates this disposition, lacking the emotional and motivational engagement virtue requires.² Virtuous action depends on cultivating a perspective on the world from which moral character emerges based on one’s accumulated experiences. Without actual perception, AI cannot be genuinely virtuous.³

Second, it is important to explore what is meant by ‘curiosity’. Arslan et al refer to curiosity as a context-seeking metric through generating clarifying questions. In the Encyclopédie, curiosity as a natural inclination is prompted by a desire to seek out more information about a novel object or concept.⁴ This desire is rooted in understanding our ignorance about said object. This perspective implies that curiosity is following an interest, one that may not correspond with a particular goal beyond the satisfaction of the desire to understand.

It is unclear whether LLMs can be curious, in part because an LLM or AI tool will not pursue a question or topic based on its own interest. AI models can learn, but the act of learning does not matter to a model.⁵ A results-focused view of so-called curiosity misses what it means to desire knowledge acquisition as an intrinsic good.

Like Arslan et al, we support designing AI to scaffold virtuous behaviour in humans; however, rather than claiming the AI itself is virtuous, we urge researchers to consider how attributing virtuousness to AI anthropomorphises them. Future research should consider adopting language that mitigates the degree to which AI models are portrayed and viewed as moral experts.