Do people trust AI medical advice as much as a GP? In a Mumbai experiment, 86.5% said they would follow the AI versus 97.3% for the doctor
Seven hundred and ninety-seven adults rated hypothetical advice labelled as AI-generated or GP-generated. Both sources were seen as similarly attentive to individual circumstances within the study's equivalence bounds, but GP advice scored higher for trust, accuracy and stated willingness to follow it. The experiment measures perceptions—not clinical correctness, real behaviour or health outcomes.
The short version
- What happened
- A professional research agency recruited 797 adults in Mumbai. Each participant evaluated four of eight hypothetical medical-advice vignettes in a partially within-person 2×2×2 design, yielding 3,188 vignette ratings.
- Who it affects
- Patients may welcome clear, immediate information from AI while still needing an accountable clinician for diagnosis, uncertainty and escalation.
- What remains unproven
- The 797-person Mumbai agency sample is not representative of India, South Asia or global populations.
Keep in mind
The 797-person Mumbai agency sample is not representative of India, South Asia or global populations.
Hypothetical vignettes measured ratings and stated intentions, not real advice-taking, clinical accuracy, behaviour or health outcomes.
Source labels and simplified messages cannot reproduce examination, follow-up, continuity, cost or accountability in a real GP consultation.
The initial three-item epistemic composite had very poor internal consistency (alpha 0.15) and was retained only for exploratory analysis.
Null effects for transparency and clinical risk are specific to the study's manipulations and should not be read as evidence that disclosure or risk never affects trust.
Source analysis · 1 linked source. Primary analysed source date: 5 November 2026.
Read the full report · 9 min The evidence, methods, implications and limitations
The answer: participants rated the AI advice favourably, but still gave GPs a clear trust advantage
Adults in a Mumbai vignette experiment did not dismiss medical advice merely because it was labelled as coming from AI. Eighty-six and a half per cent said they would follow the AI-labelled advice, and participants rated it as recognizing their individual circumstances about as well as GP-labelled advice within the study's predefined equivalence range. Yet the doctor label still mattered. GP advice was judged more trustworthy and accurate, and 97.3% said they would follow it. The gap is a reminder that receiving a usable answer and trusting the actor behind it are not the same social judgment.
The research does not show that AI medical advice is safe, correct or a substitute for primary care. Participants assessed hypothetical vignettes, not their own treatment, and the outcome was stated intention rather than observed behaviour. The advice's clinical accuracy was not independently tested as an intervention. What the study adds is a structured view of public evaluation in an Indian city: people can value AI for accessibility and clarity while still preferring a human clinician's expertise, accountability and relational care. That distinction matters for designers and health services deciding where AI should inform, explain or triage rather than act as an unaccountable medical authority.[1]
How the experiment worked: 797 people, eight vignettes and 3,188 observations
A professional research agency recruited 797 adults in Mumbai. The study used a partially within-participant 2×2×2 factorial design. Vignettes varied who supposedly produced the advice—an AI system or a general practitioner—whether the source gave higher or lower transparency about its reasoning, and whether the clinical situation was higher or lower risk. Each participant evaluated four of the eight possible combinations, producing 3,188 vignette observations. Repeated ratings from the same person were handled with mixed-effects models rather than treated as fully independent respondents.
The main outcomes were kept separate after an important measurement problem emerged. Trustworthiness was assessed directly. Recognition of the patient's individual uniqueness used two items that were strongly consistent with one another, with a reported alpha of 0.944. Accuracy and the binary intention to follow the advice were analysed separately. An initial three-item composite intended to capture epistemic trust performed poorly, with alpha 0.15, so the authors retained it only as exploratory. That decision improves transparency but also shows how difficult it is to compress trust, accuracy and feeling understood into one scale across medical scenarios.[1]
The distinction that matters
What the Mumbai experiment did—and did not—compare
Participants judged written vignettes whose source, transparency and clinical risk were varied; they did not receive treatment.
Source
AI or GP label
Tests attribution effects
Message
Higher or lower transparency
Tests disclosure framing
Scenario
Higher or lower clinical risk
Tests context
Outcome
Ratings and stated intention
Not behaviour or health
What equivalence means—and what it does not
The study did not infer 'no difference' simply because a conventional significance test failed. For perceived recognition of unique circumstances, the researchers used equivalence testing with bounds of plus or minus 0.20 standard deviations. The AI- and GP-labelled conditions fell within that operational range, supporting the narrower claim that participants did not meaningfully distinguish the two sources on those two items under the experimental setup. Equivalence is useful here because it asks whether any difference is small enough to be treated as practically negligible, given a bound chosen in advance.
That result should not be expanded into a claim that AI and GPs provide equivalent care. The measure captures whether written advice seemed to recognize an individual, not whether a clinician took a history, examined the patient, detected risk, negotiated a plan or remained accountable. A well-written vignette can sound personalized without incorporating the clinical information required for safe care. Conversely, a terse GP response in a vignette cannot reproduce the relationship and continuity of a real consultation. The experiment isolates source labels and message features precisely by simplifying the encounter; that strengthens causal interpretation of perceptions while limiting clinical realism.[1]
Why participants preferred the GP even when the AI sounded attentive
The quantitative and qualitative findings point in the same direction. Participants valued AI-labelled advice for accessibility, speed, clarity and the possibility of receiving information without delay. Those qualities can be important where appointments are scarce, travel is costly or people want a private first explanation. But GP-labelled advice carried professional expertise, accountability, safety and human relationship. A clinician can ask follow-up questions, notice uncertainty, examine a person and take responsibility for escalation. The source label therefore supplied information about the process and obligations behind the words, not just the wording itself.
This helps explain why willingness to follow was high in both conditions but still differed by almost eleven percentage points: 86.5% for AI and 97.3% for GP advice. The binary measure may also compress hesitation. A person who answers yes could still seek confirmation, delay action or choose only low-risk recommendations. The study found that transparency and scenario risk did not significantly change the continuous evaluative outcomes, but that does not mean those features never matter. Their effect may depend on what is disclosed, how severe the decision is, participants' prior experience or whether the system can explain uncertainty and route someone to care.[1]
The practical lesson is to design for escalation and accountability, not superficial imitation
For health services, the result supports bounded uses such as explaining common information, helping patients prepare questions or directing them to the appropriate level of care. It does not support replacing clinical judgment with a chatbot because users report that a vignette sounds individualized. A service should state whether content is generated, what sources and limits apply, when a clinician reviews it and how to reach urgent help. High stated acceptance makes these safeguards more important, not less: an accessible system can influence many people even when its errors or omissions are hard for them to detect.
For clinicians, the GP advantage is not simply reassurance that people prefer humans. Participants highlighted functions that systems must preserve: responsibility, contextual inquiry and relational safety. A clinician-supported tool could use AI to make information clearer while keeping escalation and accountability visible. Designers should resist making AI perform a fictional human identity or conceal its role. Trust gained through a misleading interface is fragile and ethically different from trust earned through accurate disclosure, reliable performance, appropriate uncertainty and a real route to professional review.[1]
Mumbai provides valuable evidence, but not a national or global opinion poll
The sample came from one large, highly connected Indian city and was recruited through a professional agency. Mumbai is a meaningful setting for studying digital health attitudes, but the 797 participants are not a probability sample of India or South Asia. Language, income, digital access, rural health provision, trust in institutions and experience with primary care could all alter the findings. The paper's authors are based at the National University of Singapore, and the work was supported through NUS master's research and project funds. They report no competing interests.
The study is peer reviewed and its factorial, repeated-measures design allows cleaner comparisons than a general attitude survey. Still, vignettes depend on participants imagining situations and accepting source labels. The experiment cannot reproduce anxiety, pain, cost, family influence or urgency during a real illness. It also cannot tell whether respondents understood the capabilities of current medical AI or pictured a more advanced system. Cross-cultural replication should translate and adapt the scenarios rather than assume that identical wording carries identical meanings in different health systems.[1]
What evidence would change the assessment
A stronger next step would observe real decisions with clinically validated advice and safe oversight. Researchers could compare AI-only information, GP-only consultation and clinician-plus-AI support, measuring comprehension, follow-through, appropriate escalation, errors and health outcomes. Recruitment should include rural and urban populations, multiple Indian languages, varied levels of health and digital literacy, and people with chronic as well as acute conditions. Any automated recommendation should be independently checked for clinical quality before exposure, with urgent pathways and adverse-event monitoring in place.
If transparent clinician-supported systems improved comprehension or access without increasing inappropriate self-treatment, the case for bounded deployment would strengthen. If favourable ratings collapsed when advice was wrong, uncertainty was disclosed more concretely or real costs and risks appeared, the present acceptance finding would look more fragile. The current study's most useful contribution is not a popularity contest between AI and doctors. It shows that people can see practical value in AI while reserving deeper trust for the professional relationship and accountability around a GP. Responsible systems should build around that distinction rather than trying to erase it.[1]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring primary care: compare the finding and its evidence limits.
Will patients avoid doctors who say they use AI?
A preregistered experiment with 1,030 US adults found lower ratings and appointment intentions for fictional family doctors who said they used AI. The effect is relevant to patient trust, but an advert-based intention is not a real healthcare choice.
Read the next reportExplore the full source and evidence record
What this means for people
- Patients may welcome clear, immediate information from AI while still needing an accountable clinician for diagnosis, uncertainty and escalation.
- GPs can use the finding to protect the relational and safety functions that message-only systems cannot reproduce, even when AI supports explanations.
- Health-service designers should treat high stated willingness as a reason for stronger disclosure, validation and routes to human review.
Global context
Interest in AI medical advice is growing where clinician access is uneven and smartphones are widespread, but trust is shaped by local health systems, language, regulation and experience. Mumbai's result offers a useful South Asian data point rather than a universal preference. Comparisons elsewhere should retain the distinction between an accessible information tool and an accountable clinical relationship.
What the evidence does not yet show
- The 797-person Mumbai agency sample is not representative of India, South Asia or global populations.
- Hypothetical vignettes measured ratings and stated intentions, not real advice-taking, clinical accuracy, behaviour or health outcomes.
- Source labels and simplified messages cannot reproduce examination, follow-up, continuity, cost or accountability in a real GP consultation.
- The initial three-item epistemic composite had very poor internal consistency (alpha 0.15) and was retained only for exploratory analysis.
- Null effects for transparency and clinical risk are specific to the study's manipulations and should not be read as evidence that disclosure or risk never affects trust.
What to watch next
- Field studies measuring actual follow-through, appropriate escalation and health outcomes.
- Replication across Indian languages, rural settings and different levels of digital and health literacy.
- Clinician-plus-AI comparisons rather than AI-versus-doctor framing alone.
- Independent checks of advice quality and whether trust changes after errors or concrete uncertainty disclosure.
- Governance that makes source, limits, professional review and urgent-care routes visible.
Living evidence record
Impact record IAI-1AM698O
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
6 November 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Scientific Reports published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 6 November 2026
Research question, topics and regions
How adults in Mumbai evaluate hypothetical medical advice attributed to an AI system or a general practitioner, and whether transparency or clinical risk changes those judgments
Topics: Trust · Medical advice · Public attitudes · Primary care · Transparency · Accountability
Regions: India · South Asia · Singapore · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.