What It Takes for Patients to Trust Medical AI: Key Findings from a National Conjoint Study
Overview
Artificial intelligence is rapidly expanding its footprint across clinical medicine — from radiology interpretation and sepsis monitoring to diagnostic support and clinical documentation. Yet the transformative potential of these technologies hinges on a factor that remains incompletely understood: patient trust. Without it, even the most technically sophisticated AI tool risks underutilization, avoidance, or outright rejection. A newly published, preregistered conjoint survey study in JAMA Network Open offers the most empirically rigorous characterization to date of what drives patient trust and acceptance of medical AI, with findings that carry immediate relevance for clinicians, practice administrators, and health system leaders.
Study Design and Population
Conducted between December 11, 2024, and January 1, 2025, this study employed a conjoint survey methodology — a well-validated approach for quantifying the relative weight patients assign to competing attributes in healthcare choices — among a nationally diverse sample of 3,000 English-speaking U.S. adults. The sample was demographically broad: 54.8% women, mean age 48 years, with representation across racial and ethnic groups including 12.7% Black respondents, 16.8% Hispanic respondents, and 61.9% White respondents. Income and educational diversity were similarly reflected in the sample composition.
Each participant evaluated 12 hypothetical AI-assisted clinical visits — framed around the AI-assisted diagnosis of a dermatological condition — yielding 36,000 total observations. Six key attributes were varied across visits: AI performance relative to clinicians, presence or absence of a supervising clinician, FDA approval status, Mayo Clinic certification, local hospital certification, and quality of AI training data. Respondents indicated which visit they preferred, rated their diagnostic trust on a 1-to-5 scale, and offered a brief open-ended explanation for their choice.
AI Performance: The Dominant Factor
The study's most consequential finding is the primacy of AI performance in shaping both patient choice and trust. Across all attributes tested, AI performing at or above specialist-level competence carried the greatest association with patient preference. Specifically, AI performing at the specialist level raised the probability of visit selection by 24.8% (95% CI, 23.4%–26.2%), and AI performing above the specialist level increased that probability by 32.5% (95% CI, 31.0%–33.9%). By comparison, FDA approval was associated with only an 11.1% increase in visit selection probability. The authors note that above-specialist performance was nearly three times as influential as FDA approval in shaping patient choice.
Even AI performing at the level of a general practitioner exerted an influence comparable to that of clinician presence (18.4% vs. 19.1%), underscoring that patients respond powerfully to performance signals when they are made explicit. As the authors write, "the most important factor associated with patient choice was neither any form of governance nor the presence of a clinician — it was medical AI system performance."
The Role of Clinician Oversight
Clinician presence — framed as a "human in the loop" — was the second most influential attribute in shaping patient choice, associated with an 18.4% (95% CI, 17.3%–19.5%) increase in visit selection probability. This finding aligns with prior literature documenting the central role of the clinician-patient relationship in mediating trust in healthcare encounters. Qualitative responses corroborated the quantitative results: clinician presence was the second most frequently cited rationale for visit choice, mentioned by 22.67% of respondents, trailing only AI performance (mentioned by 25.7%).
However, the authors are careful to note the limitations of this finding. Clinician oversight, while meaningful to patients, represents a resource-intensive and geographically inequitable trust mechanism. As the study observes, the availability of adequately trained clinicians is constrained, "especially in low-resource settings, for underserved populations, and in underserved specialties." Strong patient preference for a clinician in the loop may, paradoxically, limit the capacity of medical AI to extend care into the communities that could benefit most.
Governance Mechanisms: Valued, but Not Equally
Respondents expressed meaningful preferences for all tested forms of AI governance relative to no governance — a finding that affirms the baseline value patients place on oversight structures. FDA approval and Mayo Clinic certification each increased visit selection probability by 11.1%, a statistically indistinguishable effect. Local hospital certification, while still positively associated with patient choice, carried a smaller effect (7.8%; 95% CI, 6.8%–8.8%), and was significantly less influential than either federal or national-level validation (P < .001 for both comparisons).
This gradient has notable policy implications. The authors suggest that local governance, while practically indispensable for validating AI performance within specific care environments, "is the governance level most vulnerable to resource disparities and other differences between local health systems." The finding that local-level validation resonated less with patients does not diminish its clinical importance; rather, it highlights a communication and investment gap that health systems and practice leaders should consider addressing.
Training Data Transparency and Equity
A particularly instructive finding involves the disclosure of AI training data composition. Patients who were informed that an AI system was trained on a representative U.S. population dataset were significantly more likely to select that encounter (AMCE, 0.119; 95% CI, 0.106–0.131) compared to patients who received no information about training data. By contrast, disclosure that AI was trained on a dataset disproportionately composed of White, male, and wealthy individuals had no statistically significant negative effect relative to receiving no information — a nuanced finding that may reflect limited baseline awareness of the implications of non-representative training data.
These results suggest that proactive transparency about data representativeness can function as a trust-building mechanism, with particular relevance for practices serving racially and ethnically diverse patient populations. As the study's authors observe, "increasing transparency regarding these attributes may increase patient trust and consent to incorporating AI in their care."
Gender Differences in AI Trust
One demographic subgroup finding merits clinical attention. While men and women demonstrated no significant differences in how they responded to specific AI attributes — performance, governance, clinician presence, or data quality — women exhibited a meaningfully lower baseline level of trust in AI-assisted encounters overall. This finding suggests that AI adoption strategies targeting all patients uniformly may be insufficient; practices may need to consider gender-sensitive communication approaches when introducing AI-enabled clinical workflows.
Clinical and Operational Implications
The aggregate findings of this study carry several direct implications for private practice physicians and healthcare administrators navigating AI adoption decisions:
First, performance transparency matters. When patients are provided with clear, accessible information about how an AI tool performs relative to their physician or a specialist, trust increases substantially. Practices should advocate for — and vendors should be expected to provide — accessible performance benchmarks communicated in terms patients can understand.
Second, clinician engagement remains central. Positioning AI as augmenting rather than replacing physician judgment, and ensuring that clinicians remain visible participants in AI-assisted encounters, aligns with patient preferences and may substantially ease adoption resistance.
Third, governance signals are cumulative. While no single governance mechanism approached the trust impact of AI performance, the study demonstrates that FDA approval, national certifications, and local hospital endorsements each contribute positively. Layering these signals — and communicating them to patients — represents a low-cost opportunity to strengthen acceptance.
Fourth, data equity is a communicable value. Practices serving diverse communities may find that transparent disclosure of representative training data functions as a meaningful trust signal, reinforcing an organizational commitment to equitable care.
Limitations and Future Directions
The authors acknowledge several important limitations. The conjoint design assumed patient awareness of AI involvement in their care — a condition that does not reliably obtain in current clinical practice. The diagnostic scenario tested, a dermatological rash of moderate clinical complexity, may not generalize to higher-stakes or more complex conditions where trust dynamics could differ substantially. Additionally, stated preferences in survey experiments may not accurately predict real-world decision-making, and the sample, while nationally diverse, was restricted to English-speaking adults with internet access.
Future research examining combinations of governance attributes, as well as the interaction between disclosure practices and informed consent frameworks, will be necessary to translate these findings into actionable policy and implementation guidance.
Conclusions
This landmark conjoint study provides the healthcare community with an evidence-based framework for understanding what patients need in order to trust medical AI. Performance, clinician presence, governance, and training data transparency each matter — but not equally. As the authors conclude, "ensuring resource-appropriate combinations of these tools is an important step in helping AI achieve its transformative potential for health, as it is increasingly integrated into medical practice." For the private practice physician, these findings offer a roadmap: communicate performance clearly, keep clinicians visible, pursue governance layering, and lead with transparency. The patients are ready to trust AI — under the right conditions.
.png)
.png)
.png)
.png)
.png)
.png)