Patients generally support using AI to help draft responses to patient portal messages, as long as physicians review the replies before sending them and remain responsible for their content, according to a qualitative study at a large academic health system. Patients also favored clear disclosure of AI use and said the tone and length of messages should match the clinical situation.

The findings, published in JAMA Network Open, come as health systems increasingly use large language models in electronic health records to help physicians keep up with increased portal correspondence. Earlier survey findings had revealed that patients often preferred AI-drafted replies over physician-written ones.

In this follow-up study, the investigators, including lead author Kellie Owens, PhD, of the Department of Population Health, New York University Grossman School of Medicine, explored why patients felt that way and what safeguards they believe should be in place before the technology is widely adopted.

Researchers from Duke University, interviewed 40 adult patients between April and August 2025. All participants had previously completed a survey about AI-generated patient portal messages. These patients were selected to include a range of perspectives, including older adults, patients from racial and ethnic minority groups, and those who had been less satisfied with AI-generated messages in the earlier survey, resulting in a cohort of 30 women, nine men, 14 Black patients, 13 White patients, and 13 patients aged 65 years or older.

During 45- to 60-minute one-time interviews, participants reviewed sample portal message exchanges involving common situations, such as prescription refills, medication adverse effects, and potentially abnormal test results. They compared draft replies that differed in tone, length, and whether they appeared to be written by AI or a physician, then discussed their preferences for AI use and disclosure.

Survey Findings

Four main themes emerged from the interviews: Patients viewed portal messaging as a practical way to handle routine health-care needs, were comfortable with AI-drafted messages as long as physicians reviewed them, wanted message tone and length to match the clinical situation, and supported disclosing when AI was used.

Across the interviews, patients described portal messaging as a way to quickly handle routine clinical needs rather than build relationships with their physicians. Most preferred brief, direct answers and did not expect emotional support through electronic messages. Even patients who were uneasy about AI in other areas of medicine said they were more comfortable with its use in portal messaging, viewing online communication as less personal than face-to-face care.

Participants also saw AI as a practical way to help physicians manage growing workloads. Many recognized that physicians juggle numerous responsibilities and believed AI could help them respond more quickly to patient messages. However, patients consistently said they were comfortable with AI only if physicians reviewed, edited when needed, and approved every message before it was sent. They viewed physician oversight as essential for maintaining trust and accountability.

Opinions varied on the ideal style of portal messages. Some felt AI-generated responses were too long, generic, or scripted and instead preferred brief, direct replies that felt more authentic. Others liked AI's more courteous and reassuring tone, saying it showed attentiveness and invited further communication. Overall, no single style emerged as the favorite. Instead, patients preferred messages where the tone, level of detail, and length matched the purpose and urgency of the situation, regardless of whether they were drafted by AI or a physician.

Preferences also depended on the clinical situation. Most patients preferred brief responses for routine requests, such as prescription refills, but wanted longer, more reassuring messages when test results were uncertain or potentially concerning. Some said they would rather discuss serious health issues by phone or during an office visit than through the patient portal. Others felt AI could be especially helpful in more complex situations if it helped busy physicians craft more thoughtful, supportive responses.

Views also differed on what constituted an empathetic response. Many patients said empathy was less about emotional language and more about whether the response addressed their concerns and fit the clinical situation. Some felt brief messages seemed more genuine, while others preferred more reassuring language, even when they knew AI had drafted the initial response. In either case, patients said physician review was essential for trusting the message.

Patients strongly supported being told whenever AI was used to help draft portal messages, although opinions differed on how and when that information should be shared. Some preferred to be notified during clinic visits and reminded in portal messages, while others wanted disclosure with every AI-assisted reply. Most favored simple, straightforward statements explaining that AI helped draft the message and that a physician reviewed it before it was sent, rather than language promoting the technology’s benefits.

Implications and Commentary

For physicians and health systems, the findings suggest AI should be used to support, not replace, physician communication with patients. The authors recommend requiring physicians to review every AI-drafted message before it is sent, tailoring message tone and length to the clinical situation, using clear language to disclose when AI is involved, and monitoring how the technology affects patient satisfaction, communication quality, safety, equity, and physician oversight over time.

The investigators acknowledged several limitations. The study included a relatively small number of patients from a single academic health system, which may limit how broadly the findings apply. Although the researchers recruited patients with a range of backgrounds and opinions, they found no consistent differences across demographic groups. Larger studies are needed to determine whether preferences vary by age, race, education, health status, digital literacy, or previous health-care experiences.

Even so, the authors concluded that AI can improve communication with patients if it is implemented thoughtfully.

“[P]atients generally viewed AI-drafted portal messaging as acceptable and potentially beneficial when it supported timely, efficient responses to their questions,” they wrote. “However, acceptance was conditional: participants emphasized that clinicians must remain accountable through review and oversight.”

In an accompanying commentary, Teva Brender, MD, and Elizabeth Dzeng, MD, PhD, MPH, of the Department of Medicine at the University of California, San Francisco, wrote that the study adds important nuance to the growing debate over AI-generated patient portal messages by showing that patients are generally comfortable with AI when it is used transparently and under physician supervision.

Rather than viewing AI as a replacement for clinicians, the authors suggest it can help reduce administrative burden by drafting routine portal messages, as AI is increasingly being framed as a supportive administrative tool. They also note that AI could serve as a tool for improving clinicians’ communication skills, as AI could potentially help physicians craft more thoughtful responses in sensitive situations.

At the same time, the authors caution that patient attitudes toward AI are likely to change as the technology becomes more familiar, making current findings "key snapshots in time: provisional, not permanent."  For now, however, the message is clear: “disclosure and oversight remain essential.”

The bottom line, the authors wrote, is clear: “It appears that patients are not ready for clinicians to turn their in-baskets over to machines.”

DISCLOSURES: The study received no external funding. Dr. Poon reported consulting for Triomics outside the submitted work. Dr. Brender reported a grant from the UCSF Academy of Medical Educators for AI communication training research, outside the submitted work. No other authors reported conflicts of interest. 

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.