Will patients wait for a clinician to check AI medical advice?
In a survey of 1,185 US adults, 63.1% preferred waiting up to 48 hours for clinician-reviewed medical portal answers, even as majorities accepted five-minute AI responses for logistical tasks. The study tested hypothetical choices, not real behaviour or safety.
The short version
3 key points
- 1.Among 1,185 US adults, 748 people — 63.1% — preferred a health professional to review and edit an AI answer to a medical portal message, despite a wait of up to 48 hours.
- 2.Participants were more willing to choose a five-minute AI answer for logistical messages such as appointments and refills (57.4%) than for medical questions (36.9%).
- 3.The nonprobability online survey used deliberately different waits and modalities, so it cannot show how real patients would act or whether clinician review improves safety.
The answer: most accepted AI speed for logistics, not for medical advice
When the message concerned symptoms, treatment or another clinical issue, 748 of 1,185 surveyed adults preferred to wait as long as 48 hours for a doctor or nurse to review and edit an AI-generated reply. That leaves 437 participants, or 36.9%, choosing an AI answer within five minutes. The preference reversed for lower-stakes logistics: 680 participants, or 57.4%, chose the faster AI option for appointments and prescription refills.
The result suggests that people do not hold one blanket view of health AI. They distinguish between tasks and appear to value professional accountability more when a reply might influence care. For portal designers, the finding argues against quietly applying one automation rule to every inbox category simply because the technology can generate a response.
It does not prove that forty-eight-hour review is the best service model. The survey forced a trade-off between speed and oversight, and no participant received care. A real deployment could offer an immediate acknowledgement, automated routing and a later signed answer rather than either of the two options tested.[1]
The study recruited a quota-balanced but nonprobability US sample
Researchers recruited US adults able to read English through Qualtrics online panels from 17 June to 26 July 2024. Quotas balanced gender, bachelor’s-degree status and four age generations, while targeting a sample in which half identified with racial or ethnic minority groups. Of 1,237 people who completed the survey, 52 failed attention checks, leaving 1,185 for analysis.
The final sample included 599 women and 580 men; 466 participants identified as Black, 589 as White and 53 as Asian. Baby Boomers made up 389 participants, Generation X 332, Millennials 293 and Generation Z 169. These denominators are useful for interpreting subgroup estimates, but quota sampling is not the same as drawing a representative national probability sample.
The survey was conducted more than two years before publication. Interfaces, public familiarity and model behaviour changed during that interval. The data therefore describe preferences elicited in mid-2024, not a live 2026 measure of US portal users.[1]
Three portal-message categories exposed a clear risk gradient
For each of general, logistical and medical portal messages, participants chose between a reply from AI within five minutes and waiting up to 48 hours for a health professional to review and edit the AI response. AI was preferred by 600 people for general messages, 680 for logistical messages and 437 for medical messages.
The study reports a statistically significant difference across message types and followed the omnibus comparison with Bonferroni-corrected pairwise tests. It also used modified Poisson regression to examine associations with gender, race, generation, numeracy, health literacy and reported discrimination in medical settings.
Those models found, among other patterns, that women were less likely than men to accept AI for medical questions, while Black participants were more likely than White participants to accept it across all three portal categories. These are associations within this sample, not explanations. Access barriers, prior care experiences and question wording could all contribute, and small subgroups produce less precise estimates.[1]
Mental-health choices did not reduce to AI versus human
A second vignette compared a licensed therapist available in three months, an AI chatbot monitored by a licensed therapist available in one month and an AI-only chatbot available immediately. Participants rated each option separately rather than selecting only one.
The human therapist drew the highest endorsement: 757 people, or 63.9%. Therapist-monitored AI was endorsed by 595 people, or 50.2%, and AI alone by 525, or 44.3%. The ordering supports the broader finding that oversight matters, but it also shows that a monitored hybrid service may be acceptable to many people when access is constrained.
The design cannot isolate why. Modality and wait time changed together, so the survey cannot tell whether people preferred a therapist because the care was human, accepted the chatbot because it was faster, or reacted to assumptions not specified in the vignette. It also did not test crisis use, clinical severity, continuity or quality of therapeutic relationships.[1]
What health services should change now
Health systems considering generative replies should separate administrative assistance from clinical advice in policy, interface and audit logs. Scheduling, directions and refill routing may justify faster automation, while symptom interpretation, treatment changes and clinical concerns need a clearly accountable reviewer and an escalation path.
Patients should be told whether a person has read a reply before it appears in the portal. A label saying that AI was used is not equivalent to a named professional taking responsibility. Systems also need response-time targets for urgent messages so that a preference for human review does not become an excuse for unsafe delay.
Clinicians and operations teams should track both sides of the trade-off: turnaround time and the accuracy, correction and escalation of responses. Automation can shift workload rather than remove it, especially when staff must verify drafts or repair misleading answers. Patient choice should be meaningful and should not penalise people who request human review.[1]
Evidence limits and what would change the assessment
This is a cross-sectional survey of hypothetical choices. Nonrandomised answer ordering, wording effects and the use of a commercial online panel may influence the results. The options confounded response time with delivery model, and only English-reading US adults were eligible. Preferences are not evidence of safe performance or improved access.
A stronger next study would randomise wait times independently of oversight, vary clinical urgency and test whether people understand who remains accountable. Field trials could then compare real portal workflows on response time, clinician corrections, repeat contacts, treatment delays, adverse events, patient trust and unequal access across demographic groups.
Until those data exist, the careful conclusion is practical rather than universal: many people appear willing to exchange some speed for professional review when a message becomes medical. Health systems should design around that task-specific preference while measuring whether the resulting workflow is actually safer and more responsive.[1]
What this means for people
- Patients need a visible distinction between an instant automated draft and a reply a clinician has actually reviewed.
- Clinicians may face new verification work even when AI shortens drafting time, so staffing plans should measure review burden as well as message speed.
- People who already experience barriers or discrimination may make different trade-offs; services should preserve a genuine route to human review without longer or lower-quality care.
Global context
The study recruited English-reading adults in the United States. Its task-specific pattern may be relevant elsewhere, but acceptable waits, portal access, clinician shortages and trust in health institutions differ across countries and health systems, so the percentages should not be exported as global preferences.
What the evidence does not yet show
- The survey measured stated preferences in hypothetical scenarios rather than real portal behaviour or outcomes.
- Response time and human oversight changed together, preventing a clean estimate of either factor.
- Qualtrics quota sampling was not a nationally representative probability sample.
- Data were collected in June and July 2024, before publication in October 2026.
- No AI system generated medical advice in the study, and no clinician accuracy or safety benefit was tested.
What to watch next
- Randomised studies separating wait time from the presence of human review
- Real portal trials measuring corrections, delays, adverse events and staff workload
- Whether health systems let patients choose review levels without unequal access
- Replication outside the United States and in languages other than English
Living evidence record
Impact record IAI-1FGN347
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
14 October 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 JAMA Network Open 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 14 October 2026
Research question, topics and regions
How adults trade faster AI responses against delayed professional review for administrative, medical and mental-health support
Topics: Patient portals · Human oversight · Medical advice · Mental health · Health equity · Waiting times
Regions: United States · Global health systems
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.
Watch the background, then check the evidence
The Royal Society
AI-generated illustration · not a video frame
YouTube loads only when you choose. Press play in the player to start. If playback is unavailable, use “Watch on YouTube” below.
Healthcare, culture & human impact · Background explainer
Published 2026-07-01 · 44 min 28 sec
Is this the AI future we want? | The Royal Society
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Source checked 14 October 2026. Check available captions on the publisher’s video page.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
Health & Life Sciences
Can AI flag depression symptoms in adults with low haemoglobin?
A Chinese prediction study spanning 3,517 adults reported similar discrimination in two survey cohorts and a 400-person hospital cohort. It predicts a screening score, not clinical depression, and its proposed risk bands have not been prospectively tested in care.
8 min · 1 source
Health & Life Sciences
Can AI predict five-year colorectal cancer survival from pathology slides?
A retrospective preprint reports moderate discrimination across a US trial cohort and an external NHS dataset. The model is not ready to guide treatment, and its selected ensemble did not clearly beat the strongest comparator.
8 min · 1 source
Health & Life Sciences
When should AI autism-screening results reach families?
Twenty-six US participants favoured earlier support but warned against opaque EHR predictions arriving before useful next steps. The study maps implementation concerns; it does not validate a screening model.
7 min · 1 source
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.