This editorial appears in the August Issue of the American Journal of Bioethics
Epistemic humility is an important virtue for medical practitioners to show within clinical practice. Broadly construed, epistemic humility denotes a disposition to acknowledge the limitations of one’s knowledge and the corresponding willingness to share those limitations with others. Crucially, epistemic humility does not mean unwarrantedly downplaying one’s expertise or displaying false modesty but rather requires engaging with epistemic uncertainty. Within medical practice, this involves recognizing the fallibility of one’s judgments, remaining open to complementary perspectives, and fostering trust while guarding against misplaced certainty. Showing epistemic humility is also known to improve research and teamwork by encouraging responsiveness to evidence, integrating diverse knowledge traditions, and supporting skills like active listening, which help avoid epistemic injustices and enable “epistemic fluency”.
As these strands of work highlight, epistemic humility is not merely an individual attitude, but a key component of effective professional collaborations in medicine. As physicians are increasingly seen as “collaborating” with medical Artificial Intelligence (AI) systems, it is vital to consider how emerging socio-technical systems in medicine impact and could foster forms of epistemic humility conducive to clinical collaborations.
Human–AI Collaboration and Epistemic Humility
Collaboration between physicians and AI systems is widely endorsed as the most effective way to realize the potential benefits of medical AI. Human-AI collaboration is expected to offer advantages over relying solely on the AI system or solely on the human physician. Benefits include the augmentation of human skills, higher efficacy or efficiency than physicians alone and better patient outcomes. Nevertheless, some have warned against the issues of overreliance, automation bias and increasing dependence of physicians on AI systems, which may hamper effective human-AI collaboration. An authoritative presentation of algorithmic outputs may, for example, reinforce an unwarranted sense of certainty in physicians, who, without sufficient awareness of their own and the AI’s limitations, might be inclined to follow these outputs without adequately discussing relevant considerations with the patient.
In the academic literature, it is increasingly recognized that successful human–AI collaboration in medicine depends on factors that extend beyond considerations of efficiency and accuracy and are closely linked to key virtues and skills that guide medical practices. For instance, recently, it has been argued that AI needs to be designed to foster epistemic humility to know how to collaborate with these systems. As epistemic humility has long been regarded as an important intellectual virtue for human physicians and as a defining feature of the claims they advance, it is likely that AI systems that positively relate to this virtue are better adapted to medical practice. Yet, it is essential to avoid misleadingly anthropomorphizing AI systems as collaborative partners that possess certain virtues themselves. Given that AI systems cannot have knowledge, consciousness, self-awareness, and moral agency, they cannot possess the motivational and dispositional qualities that define virtues in human agents.
Technologies themselves are not moral agents, yet they can be designed to support, scaffold, or stimulate virtuous behaviors in human users. In Shannon Vallor’s account, technologies can function as virtue-conducive artifacts when they are structured in ways that encourage or enable humans to exercise those virtues. Building on her account, AI systems in medical contexts can be designed to “display” epistemic humility in ways that scaffold virtuous action in human collaborators. For example, an AI system might provide output that challenges physicians to justify their judgments, confronts them with recent empirical research and clinical guidelines, or notifies care providers if there are missing data from an individual patient to justify a specific course of action. Such design features could function as virtue-conducive affordances: they would prompt physicians to acknowledge the limits of their own knowledge, consider alternative perspectives, and engage actively with patient experiences and multidisciplinary input.
Depicting AI as displaying epistemic humility can be further substantiated by Rosalind Hursthouse’s action-oriented account of virtues. Her account suggests that virtues can provide action guidance, as virtues are not judged by the dispositions but by the actions one takes. This action-oriented account of virtue opens the conceptual space for attributing virtue-consistent patterns of behavior to entities that lack consciousness or moral agency, such as AI systems. Rather than asking whether such systems can possess epistemic virtues, the more pertinent question is whether they can be designed to instantiate or enact patterns of epistemic interaction that are consistent with those virtues. A medical AI system that systematically refrains from epistemic overreach, by avoiding unwarranted claims to authority or completeness, can thus be understood as epistemically humble in a derivative yet ethically meaningful sense.
On this view, epistemic humility functions as a collaborative and practice-oriented property that can be assessed based on externally observable actions rather than an intrinsic feature of the system. An AI system promoting or displaying epistemic humility does not merely generate accurate outputs; it structures epistemic relationships in ways that preserve the appropriate distribution of authority among clinicians, patients, and technological tools. By foregrounding the provisional and defeasible character of its contributions, such a system helps sustain the normative priority of human judgment and deliberation within clinical decision-making, without presupposing that human agents are epistemically infallible.
From the perspective of virtue-conducive design, AI systems can induce human virtues by creating virtue-conducive environments. They can be structured to prompt reflection on uncertainty, limits, and contextual factors that resist full formalization. Instead of presenting outputs as decisive and final answers, such systems should invite deliberation. Potential means of achieving this aim could be found in foregrounding uncertainty, in demanding justifications from clinicians, or in drawing attention to patient-centred considerations based on their experiential knowledge. Such “Socratic” features could help support, not diminish, clinical expertise, by helping clinicians calibrate appropriate epistemic self-trust, resist overconfidence, and remain open to competing considerations.
At the same time, the outputs of AI systems themselves should equally display epistemic humility. In an action-oriented sense of virtues, this would require AI outputs to reveal their epistemic limits by exposing the uncertainty of their predictions, their dependence on particular, and possibly limited, data sources, and the absence of relevant experiential information to avoid social misattributions. Especially in high-stakes medical contexts, epistemic humility would therefore not merely be an attractive design feature, but arguably a normative requirement for ensuring that AI contributes to, rather than destabilizes, the quality of clinical judgment. If AI systems are designed to consistently flag their limitations, they could well help to establish a collaborative environment in which epistemic authority is neither ceded to the system nor human expertise insulated from critique. Epistemic humility should therefore be considered a foundational design and interactional norm for systems intended for human-AI collaboration in medicine.
In practice, taking epistemic humility seriously when developing or using collaborative AI systems requires articulating how this virtue may be operationalized in medical care. In some instances, it may be sufficient for AI systems to display uncertainty through calibrated probability distributions or uncertainty quantification (UQ). This might be especially fitting in complex diagnostic or therapeutic settings, where clinicians must interpret available evidence to determine what ails a patient and which treatments seem most recommendable. By making uncertainty explicit, AI systems could prompt more cautious clinical reasoning here, reducing the risk of premature or overconfident conclusions.
Although this approach to medical uncertainty and the promotion of epistemic humility may be effective in contexts that can rely on ample biomedical evidence, other clinical situations may require an explicit signaling of epistemic boundaries. For example, displaying such a “scope boundary” could be needed when a medical recommendation also requires input based on a patient’s experiential knowledge and testimony. This boundary would indicate that experiential knowledge is missing from the system, emphasizing the importance of shared decision-making with the patient. an approach that is also in line with calls to guard against AI systems overreaching their warranted authority in treatment decisions.
Developing for Epistemic Humility
Fostering epistemic humility in and through medical AI could be a fruitful avenue for achieving the hoped-for benefits of human–AI collaboration. Epistemic humility also implies that medical AI intended for collaborative settings should avoid forms of algorithmic assertiveness that present outputs as authoritative and final, and instead include a designed posture of epistemic modesty that preserves the central role of human deliberation in clinical care.
Developing medical AI for epistemic humility requires careful engagement with how AI systems shape and affect humility at the human level. At the level of individual users, AI systems should be designed to promote reflective and self-critical decision-making, for instance by prompting users to engage more explicitly with their own line of thinking through “pointed questions” or uncertainty measures instead of clear-cut answers. Such design features can also support approaching human-AI disagreement as a signal of uncertainty rather than error, encouraging careful reassessment of assumptions on both sides. Beyond the individual level, it is also crucial to consider how clinician-AI collaboration reshapes epistemic humility in human-human collaborations more broadly and investigate how AI can be designed to support valuing the perspectives, distinct knowledge, experiences and needs of others. AI is, for example, envisioned to assist collaboration between different healthcare workers and geographically disparate healthcare teams. As Cajas Ordóñez et al. argue, there is an opportunity here to foster epistemic humility, as “humility also extends to interprofessional collaboration, where AI systems can serve as tools for democratizing clinical knowledge and supporting team-based care.” Building on this, epistemic humility should be understood as a leading design feature of AI systems that can contribute to, rather than threaten, the ability of human users to engage with the limitations of their knowledge and different epistemic perspectives.
In sum, while epistemic humility is a normatively attractive ideal for medical AI, its ethical and clinical value depends crucially on an appropriate pairing with epistemic competence, in what might be described as epistemic calibration within clinical practice. Diagnostic or prognostic outputs that are not proportionate to the level of medical evidence currently available cannot, even if cautiously framed, meaningfully support patient care, while overly cautious AI output in cases of very high certainty may prove equally misleading. This is particularly important in high-stakes medical settings, where both overreliance on automated outputs and their systematic dismissal can lead to harm. Epistemic humility, when grounded in demonstrable clinical performance, can help align the perceived authority of AI systems with their actual epistemic merits, supporting more appropriately calibrated trust and more reliable collaborative clinical decision-making.
Georg Starke, MD, PhD, Jojanneke Drogt, PhD, and Karin Jongsma, PhD