In 2023, we conducted a survey using the National Opinion Research Center’s AmeriSpeak Panel, a probability-based sample of US residents, with an oversampling of Black or African American and Hispanic or Latino respondents. The total sample size was 1787 participants, with 982 participants responding to a hypothetical use case related to the use of AI tools supporting the diagnosis of DR:

The outcome of interest was the perceived importance of FDA approval of the AI tool in the use case. We asked participants how true the following statement was for them: “It is important to me that the FDA has approved this AI tool” on a scale from not at all true (score=1) to very true (score=4). We generated a binary dependent variable “Importance of FDA approval,” grouping “not at all” or “somewhat true” into one category, and “fairly true” or “very true” responses into another, based on the distribution of the variable.

We then examined factors that might predict the perceived importance of FDA approval.

Specifically, respondents rated their agreement (1=not at all true to 4=very true) with statements capturing (1) comfort with the AI tool (“I am comfortable with this AI tool being used in this way”), (2) comfort with private company developers (“I would be comfortable with private companies making this AI tools such as IBM, Google, Johnson & Johnson”), (3) comfort with hospitals developing the tool (“I am comfortable with my hospital making this AI tool”), (4) comfort with patient portal companies developing the tool (“I would be comfortable with the company that makes my patient portal making this AI tool”), and (5) trust in responsible data sharing (“The organization that have my health information and share it have a good track record of using it responsibly”). These Likert scale predictors were treated as continuous variables, consistent with prior social science and health research. Treating these variables as continuous allows us to see how changes across the response levels are related to the outcome [24,25].

Political orientation was assessed as “strong/not so strong or lean Democrat,” “Do not lean,” and “strong/not so strong or lean Republican.” Indicator variables were created to evaluate categorical and ordinal variables (political orientation, sex, age, race or ethnicity, and education).

We conducted univariable and multivariable logistic regression models to identify predictors in the perceived importance of FDA approval for AI diagnostic tools using questions evaluating comfort with the AI tool and AI developers, and trust in health data sharing, political identity, and demographic factors. As a sensitivity analysis, we re-estimated models, treating Likert variable predictors as categorical variables and conducted ordinal logistic regression modeling the 4-level outcome variable. To assess the overall contribution of the categorical variables to the model, multicollinearity among predictors was assessed using variance inflation factors (VIFs). Sensitivity analyses are provided in Multimedia Appendix 1.

This study was considered exempt by the University of Michigan Institutional Review Board. Participants were compensated for their time according to standard National Opinion Research Center remuneration policies, US $4 per survey. The National Opinion Research Center obtained participants’ written informed consent. To ensure respondents privacy and confidentiality, identifying information was removed. As a result, the dataset used in the study were anonymized, and not personal identifiable information was accessible to the research team.

Among 982 respondents, 537 (54.7%) were female and 445 (45.4%) were male; 270 (27.8%) were Black, 237 (24.1%) were Hispanic, 429 (43.4%) were White, and 46 (4.7%) were other race groups (non-Hispanic).

The mean of comfort with the use of the AI tool was 2.56 (SD 0.95). Participants expressed moderate comfort with the DR AI tool being developed by hospitals (mean 2.45, SD 0.96) and private companies such as IBM, Google, and Johnson & Johnson (mean 2.34, SD 0.94), while the lowest mean values were reported for comfort with patient portal companies developing such tools (mean 2.32, SD 0.94). Participants also showed moderate agreement that organizations responsible for health data sharing have a good track record of using information responsibly (mean 2.37, SD 0.92). A majority of participants identified as strong or lean Democratic (541/982, 55.1%).

Overall, of the 982 participants, 658 (67%) stated that it was “fairly true” or “very true” that FDA approval of the AI tool was important (Table 1).

In multivariable logistic regression analysis, higher perceived importance of FDA approval was associated with greater comfort with the DR AI tool being used (odds ratio [OR] 1.44, 95% CI 1.11-1.87; P=.006), comfort with the tool being developed by private companies (OR 1.38, 95% CI 1.09-1.76; P=.008), and hospitals (OR 1.60, 95% CI 1.25-2.05; P<.001). Trust in responsible data sharing was also positively associated (OR 1.25, 95% CI 1.05-1.5; P=.01) with the outcome. Political affiliation played a significant role: lean-strong Republicans (OR 0.43, 95% CI 0.3-0.6] P<.001) and Independents (OR 0.63, 95% CI 0.42-0.96; P=.03) demonstrated lower odds of considering FDA approval important. Black American (OR 0.50, 95% CI 0.34-0.77; P<.001) and Hispanic (OR 0.57, 95% CI 0.38-0.86 P=.007) participants also showed lower odds of considering FDA approval important. Meanwhile, those with higher levels of education were more likely to consider FDA approval of AI diagnostic tools important (OR 1.64, 95% CI 1.02-2.62; P=.04; Table 2). Sensitivity analyses demonstrated substantively similar findings. Multimedia Appendix 1 shows the results of the regression model in which we treat all independent variables as categorical. Additionally, all VIF values were <3 (mean VIF 1.57, SD 0.66), indicating low multicollinearity and that predictors were empirically distinguishable.

Overall, we find that the public is supportive of FDA regulation of AI tools, and that need for regulation is not supplanted by comfort with or trust in AI, its developers, or health systems. Our findings suggest that people who are comfortable with the use of the AI tool for DR are more likely to desire FDA approval of tool, which may also suggest that people, even the most comfortable AI users, may be using the FDA approval process as a proxy for trust in the AI-CDS tool’s safety and effectiveness. This is surprising given our hypothesis that greater comfort might diminish the need for additional oversight. These findings are consistent with public expectations of having more rather than less regulation when it comes to AI and the companies that build them [12,18] and suggest that even with greater familiarity with AI tools, regulation is still a critical component of overall AI governance. This is likely to be true regardless of who develops the tool, though Pew Research shows that Americans favor more regulation of Big Tech companies (51% in 2024) [16,26]. Greater attention and effort must be given to ensure that regulatory agencies have broad and clear authority to regulate AI tools.

Additionally, protecting the legitimacy of the FDA and other regulatory agencies may also be key to public trust in AI tools. Transparency is a key FDA guiding principle [27].

Our findings suggest that hospitals and AI developers may need to seek AI tools that are approved or have gone through a reliable certification process to indicate their safety and effectiveness. Political affiliation was a statistically significant predictor of the perceived importance of FDA regulation—lean or strong republicans (OR 0.43, 95% CI 0.3-0.6; P<.001) and Independents (OR 0.63, 95% CI 0.42-0.96; P=.03) were less likely to rate it as important- highlighting the potential influence of political considerations on perception of FDA regulation.

The statistical significance of political affiliation, Lean or strong Republicans (OR 0.43, 95% CI 0.3-0.6; P<.001) and Independents (OR 0.63, 95% CI 0.42-0.96; P=.03), as a predictor of the importance of FDA regulation highlights the FDA’s inability to avoid political considerations [28].

Our findings also underscore the need to deepen our understanding of how public trust in AI tools is shaped. Rather than assuming transparency is both necessary and sufficient for trust, future research may consider whether trust may drive demand for transparency and oversight. Approval or certification processes from trusted entities, such as the FDA, could serve to signal the legitimacy of and increase public comfort with AI tools in health care. These insights point to the opportunity to further develop a conceptual framework that connects comfort with AI, perceived safety, regulatory presence, and public demand for accountability. Ultimately, identifying where that threshold of comfort lies could inform more effective strategies for regulatory engagement and public communication.

Our findings are based on DR as a use case and may not fully generalize attitudes toward FDA oversight of other AI tools, particularly new or emerging tools. Further research should explore additional clinical use cases to assess the broader applicability of these findings. Additionally, our analysis was focused on identifying associations rather than establishing causality, highlighting the need for future qualitative and longitudinal research. This is especially important considering the rapid evolution of AI health tools.

These findings suggest that the public may perceive FDA approval of AI diagnostic tools as important but also highlights the role of individual comfort with AI, trust in organizations, political views, age, and education level may influence these perceptions. As AI-CDS tools continue to evolve, understanding these perceptions can help guide the development of responsible and trustworthy health systems. These results are based on a nationally representative sample of US adults and may not generalize to other populations or health care contexts. Further research could explore equity implications and the experiences of diverse groups.