Recent media coverage of AI has centered around generative large language models and the burgeoning use of agentic AI. In health systems, AI is being increasingly implemented as a form of clinical decision support (CDS), supporting clinician judgment by estimating risk levels, interpreting images, or informing treatment planning [1-5].

It is more likely than ever that patients may encounter AI risk scores due to initiatives such as the US 21st Century Cures Act and OpenNotes [6-8]. Research on patient perspectives, however, lags behind clinician-focused studies [9]. Patients may now encounter AI-mediated information directly and therefore deserve clear, comprehensible explanations of how AI systems generate predictions that may affect their care [10].

Despite an imbalance in the focus on clinician perspectives on AI supporting clinical decision-making [9], recent studies have elicited more patient perspectives [11,12]. Our previous work indicates the importance of keeping patients informed about AI-derived risks to justify treatment decisions and promote adherence [13]. Several additional studies by our team and others have explored patient trust and reliance on decisions made by AI alone or solely by clinicians [11,12,14-20]. To date, however, there are few clinical scenarios where AI would make decisions and act independently without a clinician in the loop. Therefore, comparing differences in perceptions of AI-assisted clinical decision-making to human decision-making alone is more pertinent [21].

Before the introduction of machine learning–based AI tools in health care, clinicians often used, and also continue to use rule-based guidelines and risk calculators (RCs) to inform decision-making [22-24]. However, the RCs for CDS have shown mixed effects on patient adherence to treatment recommendations [25-28]. Therefore, we aim to understand if perceptions differ when clinicians are assisted with these rule-based tools, using RCs as an example, as opposed to the potentially more powerful but less readily explainable AI-based tools.

Finally, our previous work has also shown that trust and reliance on AI-informed decisions depend on the clinical prediction task [11,12,14,15]. Systematic reviews suggest that patients are less accepting of AI tools applied to “more serious or sensitive health conditions” [11]. While previous studies typically focus on a singular clinical decision, we explore differences across a series of clinical applications where RCs and AI may provide CDS, but that also have elements of serious or sensitive disease.

We evaluate patient perspectives on AI tools supporting CDS for 6 different clinical conditions, with attention to differences across sociodemographic groups. Our research questions included as follows:

We also summarize open-ended participant feedback regarding perspectives on RC and AI-based CDS to contextualize our quantitative findings.

We conducted a cross-sectional online survey of US-based adults between June 17, 2024, and July 26, 2024. We recruited participants via Qualtrics’ survey sampling service, which drew respondents to the closed survey from an opt-in online panel. Eligibility criteria included being at least 18 years of age, being able to read English, and providing informed consent within the Qualtrics platform before beginning the survey.

No personal information that would allow identification of participants was collected during the survey. To support comparisons across sociodemographic groups, we used quota sampling balanced on sex, education (with vs without a bachelor’s degree or higher), and generation (Baby Boomer, Generation X [Gen X], Millennial, and Generation Z [Gen Z]). We also targeted approximately half of the sample to identify as members of racial or ethnic minority groups. Participants were given monetary compensation for participating in the survey.

The Columbia University Institutional Review Board approved this study (IRB review number: AAAU3382). The design and results of the survey are reported in accordance with the CHERRIES (Checklist for Reporting Results of Internet E-Surveys) guidelines (Checklist 1) [29].

We developed survey items with input from experts in human factors, anthropology, human-centered AI, AI development, nursing, primary care, and survey design. Our team iteratively drafted and refined the items through multiple rounds of pilot testing with both experts and laypersons to assess comprehensibility and face validity. Excluding unavoidable technical terms (eg, “artificial intelligence” and clinical vocabulary), we targeted an eighth-grade reading level. The full survey battery of questions used during data collection is provided in Multimedia Appendix 1. The results of the generative AI questions that appear in the survey are not included in this paper.

After confirming age eligibility and obtaining consent, demographic questions were used to determine whether respondents met current quota needs. Those proceeding to the survey received definitions of “risk calculators” and “artificial intelligence,” and confirmed their understanding. Definitions stated:

Participants were asked about their behavioral intentions, trust, and preferences regarding medical decisions made by (1) a doctor alone, (2) a doctor supported by an RC, and (3) a doctor supported by AI. Each participant viewed six clinical scenarios (lung cancer, sepsis, depression, colon cancer, respiratory syncytial virus [RSV] vaccination, and heart failure). Participants were randomly assigned to 1 of 6 blocks, which included each decision source in 2 of the 6 scenarios and each clinical condition once, with no decision source repeated immediately on consecutive scenarios. The order of clinical conditions and decision sources was partially balanced to mitigate ordering effects. This process is illustrated in Textbox 1.

Scenarios were selected to represent contexts where AI-based CDS may be viable and to vary in seriousness and sensitivity, given prior findings that patients may be less comfortable with AI for serious or sensitive conditions [14]. Each scenario included a brief description of the condition (adapted from MedlinePlus patient education materials) [30], followed by items assessing intended behavior and trust in the recommendation (Table 1).

After completing all scenarios, participants indicated their preferred CDS approach (doctor alone, doctor + RC, or doctor + AI) for each condition. Participants were then invited to provide qualitative feedback about their concerns regarding each decision-support approach.

Before completing the survey, participants answered additional demographic measures on subjective numeracy (3-item Subjective Numeracy Scale) [31], health literacy (3-item Brief Health Literacy Screen) [32], perceived financial stability, comfort with the use of technology, and perceived discrimination in medical settings (7-item Discrimination in Medical Settings scale) [33].

We powered the study to detect differences in behavioral intentions across the 3 CDS conditions (doctor alone, doctor + RC, doctor + AI). Based on prior AI-related acceptability studies, we considered a 15% absolute difference (approximately a medium effect size; odds ratio [OR] 1.94) to be practically significant (α=.05; β=.80). Using these parameters in R (version 4.5.0; R Foundation for Statistical Computing), we estimated a minimum of 256 participants per decision-support group (n=768). We set 768 as the minimum target and allowed additional enrollment to facilitate subgroup analyses.

Table 2 describes the sociodemographic characteristics of 1185 participants who completed the survey.

For each clinical condition, we examined whether the distribution of participants who were likely vs undecided and likely vs unlikely to take the recommended action differed by recommendation source (doctor, doctor + RC, or doctor + AI) and sociodemographic variables. Figure 1 shows the percentage likely, undecided, or unlikely to complete the recommended behavior based on recommendation source for each clinical condition. When viewing doctor + AI decision support, as compared to doctor + RC (reference), participants had lower odds of reporting their behavioral intention as likely (OR 0.84, 95% CI 0.72-0.99; Table 3) as opposed to unlikely. Participants with higher numeracy and literacy had higher odds of reporting likelihood for clinical action, while Gen Z had lower odds. We also found significant differences based on clinical condition such that participants reported lower odds of behavioral intention for the heart failure, depression, and RSV scenarios than for the sepsis scenario (reference). Participants also had lower odds of being “undecided” on behavioral intentions, compared to unlikely, for the colon cancer, depression, heart failure, and RSV conditions. Those with higher literacy also had greater odds of reporting their behavioral intentions as undecided, rather than unlikely.

We then assessed differences in agreement with trust statements (eg, “I trust the doctor and AI to make a recommendation about a biopsy for lung cancer”), comparing distributions of participants agreeing vs being neutral, and agreeing vs disagreeing with the statements. Figure 2 displays results by condition. Participants had greater odds of trusting recommendations (agree vs disagree) made by their doctor alone compared to the doctor + RC (OR 1.35, 96% CI 1.13-1.61; Table 4), but lower odds of trusting the doctor + AI (OR 0.59, 95% CI 0.50-0.69; Table 4). Participants born in Gen X, those with higher numeracy, and higher literacy also had greater odds of trusting clinical recommendations, while women and those reporting higher discrimination in medical settings had lower odds. Compared to the sepsis scenario (reference), participants had greater odds of trusting recommendations regarding both cancers (lung and colon) but lower odds of trust for the depression and heart failure scenarios.

Participants also had lower odds of neutrality for feelings of trust, compared to distrust (disagree) for the doctor + AI than the doctor + RC condition. Gen X, Millennials, and those with higher literacy also had greater odds of neutrality, while those experiencing discrimination in medical settings had higher odds of distrust. Compared to the sepsis scenario (reference), participants had greater odds of neutrality in the colon cancer scenario but lower odds in the heart failure scenario.

Figure 3 shows preferences for the recommendation source by condition. A plurality (n=540-627, 46.2%‐53.6%) preferred doctor + RC in all conditions except depression, where a narrow plurality preferred doctor alone (n=492, 42.1%) over doctor + RC (n=468, 40.0%). Doctor + AI was the least preferred, with values ranging from 17.9% (n=209) to 21.1% (n=247). Multinomial logistic regression models identified significant omnibus effects for heart failure (P=.009), depression (P=.049), sepsis (P=.01), and RSV (P<.001).

The full results of the multinomial logistic regressions shown in Figure 3 are provided in the Multimedia Appendix 2. Additional data for the figures, as well as subgroup sample sizes, are provided in Multimedia Appendix 3.

Of the 1185 participants, open-ended response rates were high across conditions (doctor alone: n=1182, 99.7%; RC: n=1176, 99.2%; AI: n=1172, 98.9%). In open-ended responses, participants highlighted 4 major themes: decision quality and reliability; trust and perceptions; patient-centeredness and autonomy; and fairness, transparency, and ethics. Responses categorized as “lack of concerns/uncertainty” include explicit statements of no concern and blank or nonsubstantive responses. Table 5 provides theme definitions and frequencies by decision-support source.

Participants most frequently raised decision quality and reliability concerns about doctors (n=296, 25%), but also about AI (n=225, 19%) and RCs (n=201, 17%). Doctors were viewed as knowledgeable but fallible. One participant noted, “6 times out of 10 some doctors are right, some are wrong” (P133 on doctors).

Doctors were also sometimes viewed as influenced by external pressures. One participant noted, “They will do what the pharma companies push them to do” (P955 on doctors).

RCs were described as helpful but limited:

AI drew skepticism regarding errors and vulnerabilities, with participants providing responses such as:

Trust and perception concerns were most prominent for AI (n=263, 22%), followed by RCs (n=134, 11%), and least cited for doctors (n=75, 6%). Participants expressed ambivalence across all modalities. Regarding doctors, one participant noted, “not all doctors can be trusted for your health recommendations” (P398 on doctors).

RCs were met with skepticism and negative perceptions. One participant noted, “Its all numbers and everyone is different so I can’t see where this would be helpful” (P474 on RCs).

Opinions on AI varied. Some were optimistic. One participant noted, “AI will change the world of medicine” (P204 on AI).

On the other hand, several comments echoed rejection and mistrust. One participant noted, “I’ll never have faith in a machine!” (P403 on AI).

Participants valued patient-centeredness and autonomy across all decision support sources, with these concerns most prominently referenced for doctors (n=125, 11%), followed by RCs (n=109, 9%) and AI (n=69, 6%).

Some participants expressed concern that clinicians may not fully incorporate patient preferences or adequately involve them in decision-making.

Regarding doctors, several participants described concerns about not being sufficiently heard or included:

Others emphasized tensions between clinical recommendations and personal knowledge or preferences:

For RCs, participants again emphasized maintaining autonomy in final decisions. One participant noted, “The risk calculators are an added benefit, but in the end, I want to decide” (P741 on RCs).

AI was more often perceived as insufficiently personalized:

Concerns about ethics, overtreatment, and bias were raised across sources. Participants cited worries about data bias, privacy, and the transparency of both algorithmic and human decisions.

Some also noted potential hidden agendas and improper data use. Doctors were viewed as susceptible to bias and financial motives. One participant noted, “I worry they are more concerned with what is good for the hospital or their pocket than my health” (P463 on doctors).

RCs raised concerns about overtreatment. One participant noted, “I feel that there may be some unnecessary procedures based on the findings” (P76 on RCs).

AI elicited worries about bias and privacy. One participant noted, “Biases could come into play based on preexisting conditions and the human factor may be ignored and [I] don’t want my information being online” (P726 on AI).

A substantial proportion of responses fell into this category, though they reflected distinct underlying sentiments. Most commonly, participants explicitly stated they had no additional concerns (eg, “nothing” and “none at all”), accounting for 40.4% (n=478) of responses for doctors, 42.5% (n=500) for RCs, and 31.1% (n=365) for AI. A smaller proportion expressed uncertainty of opinion (eg, “I don’t know” and “unsure”), reflecting genuine ambivalence rather than an absence of concern (1.8% (n=21) for doctors, 4.0% (n=47) for RCs, and 3.7 % (n=43) for AI). The remainder were uninterpretable responses that were too brief, lacked detail, or did not contain full words or sentences, and, therefore, could not be reliably categorized (for doctors: n=155, 13.1%; for RCs: n=154, 13.1%; for AI: n=149, 12.7%).

Our study provides evidence regarding patient perspectives on realistic, yet hypothetical, implementations of AI-informed CDS. Prior work has not examined, within a single comparative study, patients’ reactions to common sources of CDS: a clinician working alone; a clinician using rule-based tools or guidelines (eg, RCs); and a clinician receiving AI support. A 2021 study found that patient participants were more likely to follow recommendations from a doctor alone or a doctor using AI than from an AI system making decisions independently, though this has been one of the few studies to quantitatively compare these decision-support options [21].

In our study, decision source perspectives for behavioral intentions, trust, and preference did not align. Participants reported being likely to take a clinical action (behavioral intention) more often in the doctor + risk score scenarios than in the doctor + AI scenarios. However, participants expressed the greatest trust in doctor-alone scenarios, followed by doctor + RC, and were least trusting of doctor + AI. Yet, similar to behavioral intentions, participants preferred the doctor + RC–informed decisions. This divergence between intentions, trust, and preferences may suggest that patients may want clinicians to leverage external, evidence-based tools that can synthesize complex information, yet they remain cautious about tools explicitly labeled as “AI,” although this hypothesis requires further investigation.

We found some consistent and some divergent patterns across behavioral intentions and trust related to the impact of the clinical condition; participants had lower behavioral intentions and trust in the depression (psychiatry referral) and heart failure (go home instead of being admitted) scenarios compared to the sepsis reference scenario. On the other hand, participants had lower behavioral intentions but higher trust for the colon cancer scenarios compared to sepsis. It is important to note that the sepsis scenario involved a proxy decision for a loved one. While it may be challenging to derive conclusive insights on AI use in specific clinical scenarios, our findings do underscore the continued importance of the clinical context and recommended action.

We found differences in both behavioral intentions and trust in certain sociodemographic groups; however, since these analyses were conducted across decision support conditions, they provide more insight into clinical decision-making in general, rather than computer-supported clinical decision-making. Findings reflect what may be anticipated based on prior work related to trust in medical care [38,39]. For example, those who reported experiencing greater discrimination in medical settings also reported lower trust in decision-making regardless of decision support condition. Higher self-rated health literacy and numeracy were also associated with greater trust and behavioral intentions. Future studies may benefit from exploring a single clinical scenario and assessing the interaction between different decision support conditions and sociodemographic characteristics to better understand how participant characteristics may impact perspectives on computer-supported decision support.

Qualitative responses echoed prior literature on reliability, trust, personalization, and bias [40], while adding a side-by-side critique of both computer-supported and human decision-making. Comments revealed interest in transparency in both humans (doctors being influenced by pharmaceutical or other financial incentives) and AI (lack of explanation of results). Participants described doubts and perceived deficiencies for each decision source, reinforcing the idea that all approaches have trade-offs and appropriate use cases. Qualitative findings revealed a potential misconception that AI cannot personalize care. In fact, predictive models can incorporate large volumes of patient-specific data and dynamically adjust variable weights, but this capability is not intuitive to patients. While a substantial share of participants reported no additional concerns (40.4% for clinicians, 42.5% for RCs, and 31.1% for AI), a smaller subset expressed genuine uncertainty of opinion (1.8%, 4.0%, and 3.7%, respectively). This latter group, combined with the potential misconception around AI’s personalization capabilities described above, suggests that gaps in understanding of how AI tools work persist for at least a portion of patients, even as most did not report active concerns.

Our findings offer future avenues of investigation for implementing AI in clinically realistic ways. While AI-related safety concerns must be continuously monitored and mitigated, AI can deliver benefits in predictive performance and clinical efficiency. To preserve trust, these benefits should be communicated to patients clearly and transparently.

Qualitative findings highlight the potential value of simplified, inclusive, and digestible general patient-facing education about AI. Many participants wanted to learn how AI works and how it supports their health. Comments about a lack of personalization suggest gaps in understanding AI’s capabilities and limits. Videos and tutorials that explain basic AI concepts may be an engaging starting point; they can be adapted to health care with a strong emphasis on safety and appropriate use. Although the AI decision source did not change participants’ stated likelihood of following recommended clinical actions, we observed lower trust in clinicians supported by AI for predictive tasks and stronger preferences for rule-based “risk calculators” with easily explainable outputs. Given ethical obligations and legal initiatives [6,7], it is important to disclose when AI supports treatment recommendations. To avoid eroding trust, disclosures should provide context-specific information about the tools used. This can include information such as how and when the AI adds value beyond human judgment, performance on representative populations, data provenance and uncertainty, and assurances that clinicians remain in control. In alignment with prior work, information should make system capabilities and limitations visible and support efficient correction and oversight [41]. These and previous results suggest that information should also be tailored to the clinical condition, as people value actionable, question-driven transparency over generic disclosures [42]. Implementation may benefit from reusable design patterns and templates that can be customized to the site and clinical context. Providing Fast Healthcare Interoperability Resources–based components and detailed implementation guides may support scalability across settings. Participants appeared more receptive to terms like “risk calculator” than “AI.” They also valued explainability, underscoring the importance of patient-facing explanations (eg, salient features or illustrative examples). This work should draw on decades of research in patient risk communication to avoid reinventing the wheel [43,44]. All of these approaches should be coupled with rigorous human-centered design evaluations that can assess the efficacy of these approaches, which we did not directly evaluate in our study.

Our study has several limitations that also provide implications for future work. Our quantitative findings were likely influenced by our provided definitions of RCs and AI. We had to generate our own patient-facing descriptions, as health care–centric, lay definitions of AI and related topics were not found on sites such as MedlinePlus. The informatics, risk communication, and human-computer interaction communities may support generating and coming to consensus on standardized lay-focused definitions for key AI vocabulary to mitigate this issue in the future. Given the nature of our approach, we were not able to conduct investigations regarding sociodemographic differences with the qualitative data. Future qualitative investigations may use maximum variance sampling to better tease apart potential differences.

Our survey involved sampling from an opt-in online panel, hypothetical clinical decisions, and stated behavioral intentions rather than actual clinical behavior. Future work should consider enrolling patients experiencing specific clinical conditions and measuring how different implementations of AI tools affect actual health behaviors.

Our study was also conducted only with adults in the United States. We were also unable to account for previous experiences with particular clinical conditions, which may have influenced the trust and behavioral intentions of individual respondents. Moreover, as the scenarios involved hypothetical medical decisions rather than actual care decisions, the participants’ responses could differ when compared with a real-world setting. Future work should assess similar research questions in other settings and may also benefit from understanding how patients’ lived experiences concretely affect their perspectives.

We also did not formally test for order or carryover effects across the six scenarios; while our design prevented consecutive repetition of the same recommendation source, future work could examine whether scenario position influenced responses.

We found that participants had greater behavioral intentions to take action and preferred scenarios where clinicians were supported by explainable, rule-based tools, such as RCs. Clinicians, however, remained the most trusted source for treatment recommendations, yet many participants preferred doctors using RCs—suggesting a desire for tool-supported judgment without ceding authority to AI. The nature of clinical conditions and actions also affected behavioral intentions and trust. The heterogeneity we observed, including condition-specific and sociodemographic differences, argues for context-aware deployment and co-design with diverse communities to avoid widening disparities. Together, these findings support a pragmatic path forward: use AI where it clearly improves access and efficiency, keep clinicians in the loop for consequential decisions—especially in mental health—and pair implementation with patient education, rigorous evaluation, and equitable deployment.