Patients drew a hard line between medical advice and administrative speed
Speed did not win when the message sounded medical. In a survey of 1,185 U.S. adults, 63.1% chose to wait up to 48 hours for an answer generated by AI but reviewed and edited by a doctor or nurse, rather than receive the same AI answer within five minutes with no human review. The result suggests that patients still assign a premium to clinical oversight when the question feels diagnostic, therapeutic, or otherwise medically consequential.
The preference reversed when the message became more routine. For logistical requests such as appointment scheduling and prescription refills, 57.4% wanted the faster AI response. For general messages, including nonmedical questions and administrative requests, the split was nearly even at 50.6% favoring speed. In other words, patients appear willing to trade oversight for convenience when the task is clerical, but not when the exchange touches the substance of care.
The survey tested speed against human review, not AI against human authorship
Researchers from Columbia University School of Nursing and Vanderbilt University Medical Center fielded the survey through Qualtrics online panels between June 17 and July 26, 2024. The sample was balanced on gender, education, and generation, and the design intentionally targeted a cohort in which 50% of participants identified as members of racial or ethnic minority groups. That matters because the answers did not come from a narrow pool of highly engaged portal users or a single health system’s patient base.
The question itself was carefully framed. Participants were not asked to compare an AI draft with a fully human-written reply. They chose between a fast AI answer delivered within five minutes and waiting until a health professional reviewed and edited the AI-generated response, with the delay extending up to 48 hours. This design makes the result especially relevant to health systems now trying to decide where automation ends and clinician oversight begins.
A simple operational truth follows from the data: patients were not rejecting automation outright. They were rejecting unattended automation when the question felt medically meaningful.
Racial, gender, and literacy patterns shaped willingness to accept AI
Responses were not uniform across demographic groups. Black participants, who accounted for 39.3% of the sample, were 1.4 times as likely as White participants to accept a fast AI answer to a medical question, and they were also more willing to accept AI for general and logistical messages. Women were about 24% less likely than men to accept AI for medical questions. Participants with higher health literacy were more open to AI for both general and medical questions, while those who reported more discrimination in medical settings were modestly more likely to accept AI answers to medical questions.
Generation also played a role. Generation Z participants were less likely than Baby Boomers to accept AI for logistical messages. The pattern suggests that comfort with automation is not driven by age alone and may reflect a mix of experience, trust, and perceived stakes.
These subgroup differences matter because portal policy often assumes a single patient preference. The survey points in the opposite direction: patient trust in AI is context-dependent and shaped by lived experience, especially in medical settings where prior discrimination may alter expectations about human review.
Mental health support still favored a live professional, even when slower
The survey also tested willingness to use AI in mental health support. A licensed human therapist available in three months was the most appealing option: 63.9% said they were likely to use it. Half of respondents, 50.2%, said the same of an AI chatbot monitored by a licensed therapist and available within a month. Fewer, 44.3%, were willing to use an AI chatbot alone that was available immediately.
That ordering is clinically revealing. When emotional support is the need, patients still seem to value the presence of a human clinician more than immediate access to an unsupervised chatbot. The finding fits a broader pattern in the survey: autonomy and speed carry weight, but they do not fully substitute for professional accountability when the interaction feels intimate, complex, or clinically loaded.
Portal volume is rising, so every extra layer of review has a cost
The preference for human review comes with obvious operational consequences. Patient-authored portal messages rose from 0.99 per patient per year in 2020 to 2.5 in 2025, based on Epic Cosmos data published in JAMA. At Mass General Brigham, 77.5% of portal message threads sent to primary care clinicians drew a response from the care team, but only 32.0% drew one from the clinician the patient had originally written to, based on more than 3.6 million messages.
In practical terms, human oversight can become a bottleneck. Each message routed through a clinician consumes time, delays the response, and raises the chance that automation is being used only as a first draft rather than a real efficiency gain. Prescription refills sit at the center of that tension. The Columbia survey grouped refill requests with appointment scheduling as logistical messages, the category in which most respondents favored speed. Yet renewing medications can still carry clinical risk, especially when chronic disease management, dosing accuracy, or medication interactions are in play.
The study fits a broader debate over where AI should stop
The findings arrive alongside a parallel debate about AI-generated portal messages. A 2025 Duke University Health System survey of 1,455 patient advisory committee members found respondents slightly preferred AI-drafted portal replies over clinician-written ones when they were not told who wrote them. Satisfaction dropped when respondents learned AI had generated the reply, though most still remained satisfied. The Duke data and the Columbia survey point to different questions: one asks whether patients can tolerate AI-generated communication, and the other asks when they want it delivered with or without human review.
An invited commentary accompanying the Columbia study took a pragmatic line: safety should determine the minimum level of oversight, while patient preference may guide additional review. That approach reflects the central policy dilemma. If a reply is low-risk and administrative, speed may be the correct priority. If the message could alter diagnosis, treatment, or escalation, human editing is not cosmetic; it is part of the safety net.
The study’s design leaves key questions unresolved
Several limitations keep the findings from being overread. Because response time was tied to whether a health professional reviewed the message, the survey cannot separate a preference for human oversight from a reaction to waiting. The scenarios were hypothetical, the wording may have framed answers, the sample was not a probability sample, answer options were not randomized, and only U.S. adults were surveyed. Human interaction with AI has also changed since mid-2024, the period when the data were collected.
Even with those caveats, the message for health systems is concrete. Patients may tolerate, and sometimes prefer, AI in the workflow when the task is routine. They are less willing to accept unsupervised automation when the question feels medical, and they still want a person in the loop when the stakes rise. The unanswered operational question is how much review a practice can afford before the promise of AI turns into another layer of delay.