Our survey succeeded in highlighting the barriers involved in incorporating AI-based decision-making into public health. We found that despite recognizing the social dilemma associated with AMR as a public health issue, respondents preferred a society in which both World-AI and Individual-AI are available. This indicates the complex attitudes of citizens who understand the importance of AMR but also value individual medical needs and freedom of choice in treatment. Moreover, our study highlights a division in public opinion regarding the standardization of AI diagnostics, along with significant differences in AI preferences based on gender and age. These findings provide insights into the ethical challenges and societal acceptance considerations necessary for introducing AI in the medical field.

The respondents’ opinions were divided regarding the standardization of a single AI system. Approximately half of the respondents opposed the idea that treatment guidelines should be determined by standardized AI. Furthermore, approximately half of the respondents in favor of standardization preferred Individual-AI to be the standardized AI. It can be inferred that there is a reluctance toward allowing global issues to overly influence individual diagnoses. As previous studies have indicated, current public sentiment toward AI-based medical diagnostics is a mix of resistance and anticipation^23,24,25. Considering our results, even if AI diagnostic technology is developed, it would be difficult to make World-AI a uniform standard for diagnosis through democratic methods, such as a national referendum (voting). This result also indicates that the social consensus to refrain from the excessive use of antimicrobials is not yet widespread in the general public. Note that, however, only the use of World-AI as a unified standard could lead to a fundamental solution for the AMR problem in the situation envisioned by our questionnaire. The reason for this is that even if there are only a few types of Individual-AI, the option for patients to receive diagnoses while ignoring AMR becomes available, creating a situation where patients may gravitate toward Individual-AI²². Therefore, the coexistence of World-AI and Individual-AI cannot be a fundamental solution for addressing the problem of AMR.

Here, we consider this study from the perspective of game theory. First, agreement or disagreement with standardization indicates whether respondents accept a game in which both themselves and their opponents are compelled to adopt the same strategy. When an individual can only adopt the same strategy as their opponent, he or she cannot free ride on their opponent’s cooperative behaviors. Therefore, it is understandable that those who favored standardization preferred World-AI compared to those who opposed standardization (Fig. 4). In other words, respondents who did not choose a game that allowed for free-riding desired cooperative behavior from others as well. Next, the preference for standardized AI represents the strategy preference when respondents participate in a game where they and their opponent must adopt the same strategy. Under these conditions, the selection of Individual-AI by a majority means that the preferred strategy is not cooperation but defection. This result suggests that the use of antimicrobials is influenced by game structures such as the prisoner’s dilemma game, where defection is always the optimal strategy. Therefore, if we consider the ‘repertoire of effective antimicrobials’ and ‘public health’ itself as commons (common goods/common resources), this survey demonstrated that a game structure that could lead to the tragedy of the commons due to the unregulated consumption of these shared resources (i.e., antimicrobials) exists. Interestingly, Giubilini proposed taxing certain uses of antimicrobials, particularly for treating minor and self-limiting infections in otherwise healthy individuals, as an ethically justified approach to discouraging unnecessary antimicrobial use and mitigating AMR²⁶. From a game theory perspective, such a strategy may serve as a deterrent against defection. Applying this concept to our study, it aligns with the idea of increasing World-AI adoption by setting higher consultation fees for Individual-AI than for World-AI, thereby incentivizing choices that help reduce AMR.

Interestingly, our results show significant gender and age differences in attitudes and preferences toward diagnostic AI. Individual-AI was preferred more by females than by males and by older respondents than by younger respondents, with a tendency to avoid standardization. This may reflect differences in attitudes toward AI rather than public health per se. As previous studies have shown, younger respondents and males might have a better understanding of and more experience with AI than older respondents and females^27,28. This familiarity with AI could foster a sense of trust in diagnostic AI and contribute to a more proactive stance toward AMR.

We need to carefully discuss the impact that the timing of our survey may have had on the results. To explore potential shifts in public awareness, we carried out two surveys in Japan, one before the COVID-19 pandemic and another during it. Interestingly, our findings revealed no significant changes in public perception across these periods due to the impact of the COVID-19 pandemic. This consistency in public views, as seen in both surveys, highlights the stability of public opinion regarding the issues addressed in our questionnaires. This study also underscores the challenge of raising awareness about AMR, emphasizing the need for comprehensive educational initiatives to increase the understanding of this critical public health concern. Considering the impact of the timing of the survey on the results, we may need to focus not on the onset of the COVID-19 pandemic but rather on the period when generative AI, characterized by the introduction of ChatGPT, became widespread. The adoption of generative AI makes examining whether the trustworthiness and image of AI have changed among the public worthwhile.

We also address the limitations related to the representativeness of the sample that arise from the current survey method. When using web survey companies, questionnaires are only distributed to individuals who are pre-registered as potential respondents with the company, which introduces a significant sampling bias. Additionally, these potential respondents intuitively decide whether to answer the questionnaires based on superficial aspects, such as the attractiveness of the title. Consequently, self-selection bias occurs, with individuals who have strong opinions or particular interests in the topic being more likely to participate. These biases are inevitable, even if researchers distribute the questionnaires independently, due to the limited reach available to researchers. Despite acknowledging these limitations, we recognize the utility of web survey companies due to their ability to provide a large pool of potential respondents and effectively filter out inconsistent or apathetic respondents (for details, see the Recruitment section in Materials and Methods).

Our aim is to understand the preferences of the general public, who may not be well-versed in AMR issues, rather than those of medical professionals who are already familiar with the topic. Since society is not primarily composed of healthcare professionals, it is essential to determine which preferences dominate among the general population. However, we have not assessed the variation in respondents’ knowledge. It remains unclear (1) how much knowledge respondents had about AMR before the survey, (2) how much their understanding improved through the preliminary explanation in the questionnaire, and (3) whether they fully comprehended the questionnaire content. This uncertainty leads us to hypothesize that AI preferences may differ depending on the level of AMR awareness. For future studies, combining a comprehension test on AMR with the current survey could provide valuable insights into the importance of accurate knowledge dissemination in addressing the AMR issue.

Finally, we should discuss who should be the main actor in addressing AMR. One potential objection to the significance of our study might be the suggestion that the issue of AMR could be resolved if physicians, who hold the authority to prescribe antimicrobials, manage prescriptions effectively rather than relying on the general public for prescription management. Indeed, many previous studies focusing on the excessive use of antimicrobials have predominantly regarded prescribers (i.e., physicians, hospitals, and policymakers) as the main actors in proposing solutions to AMR^{10,11,12,13,14,15,16,17,18,19,20,21}. Moreover, there are some game theory-based models and discussions that assume a social dilemma among physicians/hospitals due to AMR^{13,29,30,31,32,33}. These represent a significant number of previous studies assuming a ‘provider-driven’ society in healthcare, where antimicrobial prescriptions are decided irrespective of patient consent. However, in current societies where informed consent is established, patients themselves can make decisions about their treatment. Indeed, there are reports that pressure and expectations from patients can encourage the prescription of antimicrobials^17,21,34. In such ‘patient-driven’ societies, all citizens who could be patients should be treated as stakeholders in the AMR issue. Therefore, public opinions and sentiments about medical infrastructure become as crucial as the intentions of prescribers. Our results clearly demonstrated that public health issues, encapsulating social dilemmas such as those related to AMR, may pose a barrier to the societal adoption of even highly accurate diagnostic AI.