The use of artificial intelligence and machine learning (AI/ML) continues to expand in healthcare, with great promise for enhancing personalized clinical decision making1. As AI/ML tools become more widespread, much attention has been given to mitigating bias in algorithms to ensure they are employed fairly and transparently. However, less attention has fallen to mitigating potential bias among AI’s human users. As automated systems become more sophisticated in their capacity to predict, screen for, or diagnose disease, the temptation to rely on them in clinical decision making will increase2. However, factors that influence user reliance on AI are poorly understood, and healthcare professionals lack guidelines about the role that AI should play in their decision making. We argue for a more systematic approach to identifying the existence and impacts of user biases while using AI tools and their effects on clinical decision making and patient outcomes. Specifically, we call for greater empirical research into how to mitigate biases with anticipated negative outcomes through the use of embedded interface design features, drawing on insights from decision science and behavioral economics, to nudge users towards more critical and reflective decision making using AI tools.

Expand notions of user testing

Recognizing the potential harms of overreliance on AI systems in the context of high stakes decision making, regulators and policymakers seem to endorse keeping humans “in the loop” and focus their action plans and recommendations on improving the safety of AI/ML systems such as through enhanced computational accuracy3,4,5. Meanwhile, developers are innovating new ways of addressing trustworthiness, accountability, and explainability of “black box” AI/ML that involves deep learning or neural nets with significant interpretability limitations6,7. These goals appear to be particularly important when using AI/ML in clinical decision making, not only because the costs of misclassifications and potential harm to patients are high but also because undue skepticism or lack of trust can reduce stakeholders’ adoption of promising new AI technologies and inhibit their use and availability outside of experimental settings.

One of us (SG in Babic et al.8), however, recently warned healthcare professionals to be wary of explanations that are presented to them for black box AI/ML models.

Explainable AI/ML … offers post hoc algorithmically generated rationales of black box predictions, which are not necessarily the actual reasons behind those predictions or related causally to them. Accordingly, the apparent advantage of explainability is a “fool’s gold” because post hoc rationalizations of a black box are unlikely to contribute to our understanding of its inner workings. Instead, we are likely left with the false impression that we understand it better.”

Consequently, instead of focusing on explainability as a strict condition for AI/ML in healthcare, regulators like the U.S. Food and Drug Administration (FDA) should focus more holistically on those aspects of AI/ML systems that directly bear on their safety and effectiveness—especially, how these systems perform in the hands of their intended users. While the FDA recently published its final guidance explicitly recognizing the risks of automation bias9 and is working on a new regulatory framework for modifications to AI/ML-based software as a medical device (i.e., software that is itself classified as a medical device under section 201(h)(1) of the U.S. Federal Food, Drug, and Cosmetic Act10), Babic et al. argue that regulators like the FDA should also, at least in some cases, emphasize well-designed clinical trials to test human factors and other outcomes of using AI in real-world settings. Gerke et al.11,12 similarly argue that more algorithmic tools must be prospectively tested to understand their performance across a variety of procedural contexts that mirror their intended use settings and human-AI interactions. The type of user testing these scholars are suggesting goes beyond the typical usability and acceptability testing that characterizes the pipeline from beta to a more finalized version of an AI tool. That type of testing is most often done heuristically13, using a small set of evaluators to examine the interface and judge its compliance with relevant usability principles (e.g., interpretability, perceived utility, navigability, satisfaction with use, etc.). While these metrics are often useful for gauging proximate user experiences (i.e., “UX” testing) with a tool’s interface, a deeper level of user testing is needed14 to help identify and address potential sources of “emergent” or “contextual” bias15 that arise due to mismatches between a product’s design and the characteristics of its users, use cases or use settings. These mismatches may be more difficult to predict and account for in the case of AI tools than for traditional medical devices or pharmaceuticals whose performance is less contingent on user interactions and interpretations12, or whose adaptive algorithms continuously change16. Mitigating these mismatches can only be achieved by broadening our notion of user testing beyond its current focus on AI performance metrics and proximate usability to examine human and systemic factors shaping how AI systems are applied in practice17,18 by imperfect users in imperfect settings. Further, testing does not have to be limited to simply observing how individuals in various contexts interact with AI tools; we can also test how best to shape those interactions using existing insights from the behavioral sciences, as we discuss below.

Trust in the eye of the (imperfect) beholder

At this stage in the history of human-machine relations, nearly everyone is an imperfect user of AI. By this, we mean imperfectly rational: our interpretations and integration of information into decision making, including insights derived from AI, are susceptible to well-documented forms of bias19,20. Not all biases, however, are equally salient or relevant to the safe, effective, and responsible use of AI. From both legal and ethical perspectives, the most important cognitive biases are those that impact the extent to which humans rely on AI in their decision making in ways that introduce risk. Reliance falls along a spectrum of utter rejection or skepticism of AI on one end to “blind” overreliance or acceptance of AI-derived conclusions on the other. Both types of error can have negative impacts on patient outcomes, with underreliance potentially leading to errors of omission and overreliance on errors of commission.

Where clinical decision makers fall along this spectrum depends on how much they trust an AI system. Literature from anthropology and developmental psychology documents findings that human trust is influenced by how other people behave in contexts of reciprocity and exchange21, not only of goods and services but also attachment behaviors22,23 (e.g., affection, nurturance). Loyalty24, integrity25, and competence26 play important roles in human-human trust, increasingly conceptualized as an evolved capacity to help us navigate complex social dynamics and to mitigate personal risk by understanding which entities and objects can be trusted under which contingencies27,28,29. While we know a great deal about trust in human relationships, we are just beginning to understand how and in what circumstances humans trust machines. Literature on human-machine interactions, or “human factors” research, has existed for decades in other domains, including military, aerospace, and robotics; but only within the last decade have questions surrounding human interactions with autonomous systems (e.g., automation bias) begun to animate the field of AI broadly, and AI ethics in particular2,11.

Impacts of uncertainty and urgency on decision quality

Trust plays a particularly critical role when decisions are made in contexts of uncertainty. Uncertainty, of course, is a central feature of most clinical decision making, particularly for conditions (e.g., COVID-1930) or treatments (e.g., deep brain stimulation31 or gene therapies32) that lack a long history of observed outcomes. As Wang and Busemeyer (2021)33 describe, “uncertain” choice situations can be distinguished from “risky” ones in that risky decisions have a range of outcomes with known odds or probabilities. If you flip a coin, we know we have a 50% chance to land on heads. However, to bet on heads comes with a high level of risk, specifically, a 50% chance of losing. Uncertain decision-making scenarios, on the other hand, have no well-known or agreed-upon outcome probabilities. This also makes uncertain decision making contexts risky, but those risks are not sufficiently known to the extent that permits rational decision making. In information-scarce contexts, critical decisions are by necessity made using imperfect reasoning or the use of “gap-filling heuristics” that can lead to several predictable cognitive biases20. Individuals might defer to an authority figure (messenger bias34, authority bias35); they may look to see what others are doing (“bandwagon” and social norm effects35,36); or may make affective forecasting errors, projecting current emotional states onto one’s future self37. The perceived or actual urgency of clinical decisions can add further biases, like ambiguity aversion (preference for known versus unknown risks38) or deferral to the status quo or default39, and loss aversion (weighing losses more heavily than gains of the same magnitude40). These biases are intended to mitigate risks of the unknown when fast decisions must be made, but they do not always get us closer to arriving at the “best” course of action if all possible information were available.

Reducing or exacerbating uncertainty

One of AI’s most compelling advantages for healthcare is to reduce this uncertainty—for example, by calculating a personalized estimate that a patient’s condition will worsen after X amount of time or will enjoy a survival benefit of Y number of years post-intervention. However, whether AI successfully contributes to reducing uncertainty still depends to a large extent on how estimates are interpreted and acted upon. A small number of studies examining decisional biases when using AI have identified that physicians across expertise levels often fail to dismiss inaccurate advice generated by computerized systems (automation bias41,42,43,44,45), but as well as by humans, indicating that people are generally susceptible to suggestions. The tendency to follow even bad advice appears to be even more prevalent among participants with less domain expertise46,47. Receiving such advice from AI systems can raise further dangers by potentially engaging other cognitive biases such as anchoring effects and confirmatory bias, in which users are primed towards a certain perspective and disproportionately orient their attention to information that confirms it48. Other studies have found that participants are averse to following algorithmic advice when making final decisions (algorithmic bias)49,50,51, but this result is inconsistent with other studies, which show people sometimes prefer algorithmic to human judgment46,47,52.

Given the diversity of cognitive biases and contingencies under which they are likely to emerge, further systematic research is needed to document which salient factors shape how we integrate AI into decisions and how best to calibrate trust so that it matches what AI systems can actually do (e.g., predict something with a given degree of probability and accuracy). In robotics, poor “trust calibration” between humans and machines is viewed as a core vulnerability and key predictor of performance breakdown53,54. Likewise, putting AI in the hands of users without systematically measuring, controlling for, or otherwise trying to calibrate trust and reliance likely exacerbates rather than reduces the already high levels of uncertainty that characterize these decision-making contexts, with potentially grievous consequences.

The uncertain role of AI in clinical decision making

The current push55,56,57 to enhance healthcare professionals’ literacy in AI/ML highlights a need to replace idiosyncratic variation with informed reasoning about the role that AI should play in clinical decision making. However, it is hard to know what kind of guidance healthcare professionals should receive when so few empirical conclusions have been drawn about how AI is or should be used in clinical (or any) decision making. Taking lessons from algorithmic tools that have been shown to reproduce negative societal biases in predicting factors like criminal recidivism58, health status and insurability1, and disease (e.g., skin cancer) risk59, many scholars argue60,61 that AI tools should not replace any decisions that are considered “high stakes”—those with significant health or justice-related impacts. In the healthcare setting, some experts recommend that even AI with a well-demonstrated capacity to autonomously identify and diagnose disease should be confirmed with human-led testing62,63. Similar conclusions have been made about autonomous weapons systems (AWS) in military64 and maritime (e.g., unmanned shipping65) applications, with ongoing debates about whether to keep humans “in” the loop or “on” the loop, the latter suggesting that humans may not need to take an active role in decision making but can (and should) still intervene or be able to appeal to AI inferences when their conclusions contradict those of the AWS (if caught in time).

If we agree that humans should still be “in” or “on” the loop, how should one expect healthcare professionals to react to AI-derived information? The recommendation to proceed with caution, while warranted, seems too broad to fit the decisional needs of physicians engaging powerful AI to inform complex medical decisions. There is growing agreement that proficiency in AI (including its shortcomings related to bias, transparency, and liability) should be part of medical education, with suggestions that medical students must acquire sufficient knowledge of data science, biostatistics, computational science, and even health AI ethics66 to ensure they can, among other things, separate “information from hype” and critically evaluate AI systems57,67. Others68 have argued that learning effective debiasing strategies and cultivating awareness of how heuristics can impact clinical decision making should be prioritized in all stages of medical education. However, it remains unclear which biases healthcare providers should be made most aware of; whether providers should be responsible for being aware of their own biases, or whether bias mitigation may (or should) be embedded in standardized processes for implementing AI tools in clinical decision making or in the design of the technologies themselves.

Enhancing decision quality by design

While it is likely true that physicians will increasingly need to learn how to responsibly use AI to keep pace with clinical innovations, other complementary approaches should also be explored. One promising option is to support physicians in their likelihood to demonstrate the specific characteristics we value in clinical decision making by embedding bias mitigation techniques into the very design features of our AI systems and user interfaces. This notion builds on longstanding work in computing ethics69,70 and is known by various terms, including Value-Sensitive Design (VSD71), Values @ Play72, reflective design73, adversarial design74, and critical technical practice75, and was originally pioneered by Friedman and Nissenbaum76,77 in the 1990s as a way to encourage a reflective, iterative process for shaping human-computer interactions that prioritize trust and user welfare. A great deal of variation remains in how VSD is carried out, but the centrally motivating assumption behind this approach is that reflective design approaches can help to mitigate user biases for more favorable outcomes. Following the three main stages of VSD would entail identifying the range and diversity of stakeholder values and how best to balance them towards an articulated goal (conceptual), observing impacts of given values and practices on relevant outcomes (empirical), and devising technical specifications to design systems that reflect or help to shape the use of a system to align with stakeholders’ values (technical). An example would be to design interactive web browser cookie management systems to reflect principles of privacy, voluntariness, and right to disclosure71. Scholars have extended a fourth and ongoing step of life-cycle monitoring and evaluation to VSD for AI specifically, given the often unforeseeable impacts and adaptive nature of AI tools14,78.

Building on these approaches, we argue that a VSD approach could not only help to embed values into the design of AI tools but also to actively and strategically influence (nudge) users to engage in more ethical and critical reflection in their use of such tools. Such an approach requires critical engagement with the ethics of nudging in health decisions as well as identification of the range of target values one wants physicians to demonstrate in decision making. Nudging is a form of libertarian paternalism in which decisions are actively shaped through strategies such as information framing, structuring incentives, and other means to enhance the uptake of certain behaviors79. While evidence for the efficacy of this approach dates back nearly two decades80, nudging tactics have shown to be effective, for example, during the COVID-19 pandemic to encourage compliance with public health-promoting behaviors, such as handwashing and social distancing81. Though not without its critiques (e.g., that it can be a form of manipulation82,83), a central rationale of nudging is to preserve individual choice while guiding people toward behaviors with population-level benefits84. However, determining who gets to decide which values are engaged in service of making “good” decisions when using an AI tool is complex and should draw on perspectives from multiple, diverse stakeholders, not just those of developers designing these systems. The Hippocratic Oath establishes a fundamental criterion that physicians’ decisions should be in service of what they believe to be a patients’ best interests. Additional criteria come from a rich literature on decision making and clinical decision support85, suggesting that “quality” decisions are those that are informed and generate positive outcomes that are congruent with a patient’s values. Other target values, such as decisional autonomy82, are likely to be relevant, and it should be noted that salient target values may shift depending on the nature of the AI tool or the ethical issues raised by its intended users or use contexts. For example, an AI tool designed to predict and prevent onset of psychiatric illness in adolescents raises a particular set of target values in decision making (e.g., decisional autonomy, patients’ right to an open future) while a tool to identify presence and prognosis of lung cancer in adults may raise others (e.g., avoidance of negative emotional reactions, actionability considerations, patients’ right to not know). Research is needed to elucidate which target values for “quality” decision making are most salient in which clinical scenarios.

AI interfaces that encourage critical reflection

One target value that is likely to be relevant in all clinical decision making involving AI is the need to promote reflexivity in decision making in order to avoid the potential negative consequences of overreliance on AI. A growing literature1,86 demonstrating the potentially deleterious effects of overreliance on AI algorithms highlights the importance of reflexivity and deliberation as guiding principles for AI deployment. These explorations and observations thus inform the conceptual and empirical stages of the VSD approach, leaving the technical challenge of designing interfaces that will help to shape the deliberative and reflexive use of AI systems in ways that align with users’ interests. Research has demonstrated that the ways in which information is presented can influence users’ likelihood of engaging in reflective or critical thought. For example, a study by Zhang et al.87 employed a simple interface nudge to encourage reflection by asking participants to answer brief questions clarifying their own opinions versus what they considered to be reasons driving alternative perspectives. Weinmann88 developed an online interface with similar questions to enhance “deliberation within” by asking questions that encouraged reasoning about alternative perspectives. Other research by Harbach et al.89 demonstrates the effectiveness of using interface design elements to inspire reflection by illustrating consequences of user choices (e.g., reminding users of the potential impacts on selecting certain user privacy parameters). Menon et al.90 similarly explored how modifying “interface nudges” in relation to specifically targeted cognitive biases (e.g., anchoring and social desirability effects) influenced user deliberation and responses. These studies highlight how strategic interface design can help to enhance reflection and reduce passive reception of information.

For example specific to AI system interfaces, design elements might vary according to stakeholder type. An interface designed to reduce physicians’ overreliance on an AI model estimating a patient’s 1-year survival post-intervention might include brief questions or a checklist encouraging physicians to document which other clinical, psychosocial, or environmental factors or additional expert opinions they have consulted in order to corroborate (or challenge) the AI’s estimate. Complementarily, a patient-facing interface for the same tool may contextualize the numerical survival estimate within a more holistic values clarification exercise asking patients to circle one or more treatment goals influencing their decisions, encouraging reflective, value-based decision making. Building in such reflexivity metrics could not only help to nudge users away from overreliance on AI tools but also evaluate impacts on clinical decision making in practice, both within and beyond clinical trial contexts.

However, interfaces are not the only tools available with this capacity. Conceptualizing how an AI system might fit into clinical flow in ways that encourage deliberation among clinical teams may also help to reduce potential for overreliance91. Situational and logistical factors could be considered, such as setting (e.g., the collective use of an AI tool during medical review board vs. individually in a physician’s office), timing (before or after treatment candidacy), and information access (direct-to-patient versus physician-privileged communication of results). Integration of AI with other existing clinical technologies may also alter outcomes of using AI tools by broadening information that is integrated into decision making92. Organizational aspects may include training, supervision, handover, and information flow across members of the clinical team91.

These insights discussed above represent only the tip of the iceberg of factors that may potentially be coordinated to positively influence decision quality and outcomes using AI. They have been identified and often widely discussed in fields as diverse as decision science, behavioral economics, human factors, psychology, political science, robotics, and others. However, few of these insights have yet been integrated into AI systems design or systematically tested in clinical trials to proactively shape how AI is used.


We echo others’ calls that before AI tools are “released into the wild,” we must better understand their outcomes and impacts in the hands of imperfect human actors by testing at least some of them according to a risk-based approach in clinical trials that reflect their intended use settings. We advance this proposal by drawing attention to the need to empirically identify and test how specific user biases and decision contexts shape how AI tools are used in practice and influence patient outcomes. We propose that VSD can be used to strategize human-machine interfaces in ways that encourage critical reflection, mitigate bias, and reduce overreliance on AI systems in clinical decision making. We believe this approach can help to reduce some of the burdens on physicians to figure out on their own (with only basic training or knowledge about AI) the optimal role of AI tools in decision making by embedding a degree of bias mitigation directly into AI systems and interfaces.