The research team surveyed 1,185 adults based in the United States, asking them how they felt about the use of AI within secure medical portals, the encrypted messaging environments where patients routinely contact their care teams with questions, requests, and concerns. These portals have become a central artery of modern outpatient care, and health systems have increasingly looked to AI tools to help manage the sheer volume of messages that flows through them. The Columbia study examined a specific and practical question at the heart of that trend: how quickly do patients want AI to act, and how much human oversight do they demand before an AI-assisted response reaches them?

The answer, the researchers found, depends heavily on what kind of help a patient is seeking. For logistical requests, a majority of respondents, 57 percent, said they preferred fast AI-generated responses. These are the routine transactions of clinical administration, such as appointment scheduling and prescription refills, where the information needed is often factual, standardized, and unlikely to require nuanced medical judgment. In these lower-stakes situations, patients appeared willing to trade the involvement of a clinician for the convenience of an immediate, automated reply.

Medical questions told a very different story. Two-thirds of participants said they would rather wait up to 48 hours for a doctor or nurse to review and edit an AI response to a medical question than receive an unreviewed AI answer right away. That preference for delayed but human-verified information represents a striking rebuke to the assumption that patients value speed above all else. When the content of a message touches on diagnosis, treatment, or clinical guidance, the survey suggests, people want the assurance that a trained professional has read the words before they do, even if that assurance costs them two days of waiting.

The study’s title, Public Perceptions of Timeliness for AI Use and Human Oversight, captures this central tension between immediacy and verification. Lead author Natalie Benda, PhD, an assistant professor at Columbia Nursing, said the results confirmed what the team had suspected but could now support with data. When people have a medical question, she noted, they want to hear from their doctor, not AI, even if AI may give an immediate answer. Her co-authors included Uday Suresh, Megan E. Salwei, and Laurie Lovett Novak of Vanderbilt University Medical Center; Afra Shamnath and Meghan Reading Turchioe of Columbia University School of Nursing; Isabella Buitron of the Columbia University Mailman School of Public Health; and Xuhai (Orson) Xu of Columbia’s Department of Biomedical Informatics.

Mental health support emerged as a particularly revealing test case in the survey. Participants were presented with three options for obtaining mental health care, each differing in the balance between human involvement and availability. A licensed human therapist with a three-month wait drew the strongest support, at 64 percent. A therapist-monitored AI chatbot available within one month was acceptable to 50 percent of respondents. An AI-only chatbot available immediately, the fastest and most automated option, was the least popular, at 44 percent. The gradient is instructive: each step that reduced human oversight or increased automation corresponded to a measurable drop in acceptance, even as each step also shortened the wait for care.

That pattern matters because mental health care is one of the areas where AI chatbots are being deployed most aggressively, often justified by the severe shortage of therapists and the long delays patients face in securing appointments. The survey data suggest that patients themselves are acutely aware of this trade-off. Many would apparently rather endure a three-month wait for a licensed human therapist than receive instant support from a system with no human involvement at all. At the same time, the fact that half of respondents found a therapist-monitored chatbot acceptable indicates a meaningful openness to hybrid models, in which AI handles day-to-day interaction while a clinician supervises the overall course of care.

The study also revealed that attitudes toward AI in medical settings are not uniform across the population. Benda noted that people’s opinions differed across characteristics such as race, health literacy, gender, and experiences of discrimination in medical settings. One of the most notable findings was that individuals who reported having been discriminated against in a medical setting were more open to AI responses to medical questions. That counterintuitive result complicates any simple narrative about technology acceptance. For some patients, the prospect of an AI-mediated interaction may feel less fraught than an encounter with a health system in which they have previously felt mistreated, suggesting that the drivers of trust in AI are entangled with broader experiences of inequity in care.

The authors argue that these findings carry a direct implication for how health systems design and deploy AI tools: patients should be allowed to make decisions about how and when AI is used in their medical care. Rather than imposing a single policy across all message types, the study suggests, portals and clinics may be more successful if they let patients choose, reserving rapid AI responses for administrative tasks while routing medical questions through clinician review. The data indicate that patients may be receptive to AI in lower-stakes situations but continue to expect clinician involvement for higher-stakes medical decisions and guidance. Offering that choice could preserve the efficiency gains that motivate AI adoption while respecting the trust conditions on which patient acceptance depends.

As health systems weigh investments in AI-powered messaging, triage, and patient communication, the Columbia survey provides a quantitative anchor for those decisions. It quantifies the willingness of a large sample of U.S. adults to wait for human review, the tasks for which they will accept automation, and the demographic fault lines along which acceptance varies. The message from the 1,185 respondents is consistent and pragmatic: AI can handle the scheduling and the refills, but when the question turns to health and the answer carries clinical weight, patients want a doctor or nurse to read it first, even if that means waiting an extra 48 hours.

Subject of Research: Public perceptions of AI-generated medical responses and human oversight in patient portals

Article Title: Most patients won't endorse AI medical responses without a clinician's review, Columbia Nursing study in JAMA Network Open finds

Article References: Most patients won't endorse AI medical responses without a clinician's review, Columbia Nursing study in JAMA Network Open finds. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: artificial intelligence, patient portals, clinician oversight, JAMA Network Open, Columbia University School of Nursing, health care AI, patient trust, mental health chatbots, medical messaging, health literacy, health care discrimination, survey research