AI Advice Earns Interest, but Doctors Still Hold the Trust Advantage in Mumbai
A survey experiment in Mumbai, India, found that AI-generated medical advice can be received positively even when it is not seen as a substitute for physician judgment. Across 797 participants and 3,188 vignette ratings, advice framed as coming from a general practitioner consistently outperformed AI on trust, perceived accuracy, and stated willingness to follow the recommendation. The gap did not erase interest in AI. It did, however, show that favorable impressions and clinical reliance are not the same thing.
One result stood out sharply: participants were almost as willing to follow AI advice as they were to appraise it favorably, but they still placed greater confidence in the human clinician.
What the 2 × 2 × 2 Design Revealed
The study used a mixed-methods, partially within-subject factorial design with three moving parts: source of advice, level of transparency, and level of clinical risk. Each participant evaluated four vignettes chosen at random from eight combinations. The structure allowed the investigators to ask whether people judge advice differently when it is attributed to AI rather than a general practitioner, and whether that judgment shifts when the scenario is more transparent or more clinically risky.
The core comparisons were straightforward. GP-generated advice was rated as more trustworthy and more accurate than AI-generated advice, and it drew stronger stated intention to follow the advice. Yet the perceived recognition of individual uniqueness did not differ significantly between the two sources, and equivalence was supported within the study’s operational ±0.20 standard deviation bounds. In other words, participants did not clearly see AI as less able to recognize them as individuals, even while they still preferred the doctor on the measures that matter most for confidence and clinical reliance.
The evaluative pattern was not fragile. It held across a relatively large number of vignette judgments and survived mixed-effects modeling that accounted for repeated measures. The only epistemic composite originally created for the study performed poorly, with internal consistency too low to support primary interpretation, so it was set aside except for exploratory use. The stronger signal came from the separate trustworthiness and uniqueness-recognition measures, where the AI-versus-GP contrast was easier to see and easier to interpret.
High Willingness to Follow AI Did Not Mean Equality
The behavioral-intention data deserve special attention. Stated willingness to follow advice remained high in both groups, reaching 86.5% for AI-generated advice and 97.3% for GP-generated advice. That is a meaningful difference, but not a total rejection of AI. Participants were willing to act on machine-generated guidance far more often than critics of health AI might expect, especially given that the scenarios were framed as medical advice rather than general information.
The practical implication is clear: AI advice does not need to win full trust to gain use. Accessibility, clarity, and speed may be enough to secure attention, especially in settings where physician access is limited or delayed. Still, the higher willingness attached to GP advice suggests that people are not collapsing these two sources into the same category. The doctor remains the safer choice in the mind of the respondent, especially when the question is not whether the advice sounds reasonable, but whether it deserves reliance.
Transparency and Clinical Risk Had Little Measurable Effect
The manipulation of transparency and clinical risk did not significantly shift the continuous evaluative outcomes. That finding matters because it narrows the explanation for source preference. People in this sample were not simply responding to whether the advice was presented with more explanation or whether the vignette implied greater medical seriousness. Their judgments were driven more strongly by who was giving the advice than by those contextual details.
This pattern suggests a durable source effect. Even when the content of the advice was held within a controlled experimental structure, participants continued to attribute greater trust and accuracy to the human clinician. The source label carried weight on its own.
Why the Qualitative Data Point in Two Directions
The open-ended responses added texture to the numerical findings. AI-generated advice was appreciated for accessibility and informational clarity. Participants seemed to value the speed and convenience implied by an automated system that could answer questions without delay. That is not a trivial benefit in a system where time, geography, and clinician availability can all limit access to care.
GP-generated advice, by contrast, was linked to expertise, accountability, safety, and relational care. Those terms are not interchangeable. Expertise speaks to knowledge, accountability to responsibility, safety to risk management, and relational care to the human dimension of treatment. In a medical context, each of those features can influence whether advice feels actionable or merely informative. The qualitative themes show why AI can be liked without being fully trusted: it may be useful, but it still lacks the social and ethical weight attached to a clinician’s judgment.
What This Means for Medical AI in India
The findings fit a broader reality of clinical adoption. Patients and the public do not evaluate AI as a single, monolithic tool. They weigh convenience, clarity, and novelty alongside questions of responsibility and safety. In this study, AI was not dismissed. It was rated favorably enough to suggest real openness to machine-generated health information. Even so, the preference hierarchy remained intact, with physicians still occupying the more trusted position.
For digital health developers and clinicians, the message is not that AI must replace the doctor to matter. It is that AI may function better as a complementary source of information than as a stand-in for professional medical advice. That framing aligns with the behavioral data and the qualitative responses alike. People may welcome the efficiency of a system that can explain, summarize, and guide, while still reserving final confidence for a trained clinician.
Limits That Shape the Interpretation
The study was conducted in Mumbai and may not generalize cleanly to other regions, health systems, or cultural settings. The use of vignette-based judgments also means the findings capture stated preferences rather than real-world clinical behavior. Participants were reacting to hypothetical advice, not deciding in a live encounter with symptoms, uncertainty, or urgency.
Another caution sits in the measure set itself. One constructed epistemic composite did not perform well psychometrically, which is why the more reliable single measures carried the main analytic load. Future work will need to test whether the same source preference appears in different populations, across different kinds of medical decisions, and under conditions where the personal stakes are higher than a survey vignette can capture. The present study makes one point unmistakable: AI can win approval, but it has not yet won parity with the trusted doctor.