Doxiverse Staff - July 2026
A qualitative study published July 7 in JAMA Network Open found that patients are generally comfortable with AI-drafted replies to their patient portal messages, but that comfort is conditional: patients consistently said a clinician must review, edit, and take responsibility for the message before it's sent, and that they want to be told AI was involved.
Why it matters: As more health systems deploy generative AI to help manage the growing volume of patient portal messages, this study offers patient-sourced guardrails, not just "should we use AI," but specifically what conditions (mandatory review, disclosure, message-length sensitivity to clinical stakes) patients say are necessary to maintain trust in the tool.
What Happened
Researchers affiliated with Duke University Health System and NYU Grossman School of Medicine conducted 45- to 60-minute videoconference interviews with 40 adult patients recruited from a single large academic health system, who had previously completed a related survey on AI-drafted portal messaging. The sample skewed toward women (75%) and included 14 African American participants, 13 White participants, and 13 patients aged 65 or older; the researchers deliberately oversampled racially and ethnically minoritized patients, older patients, and patients who had reported lower satisfaction with AI-drafted messages in the prior survey, specifically to capture a wider range of views rather than to produce a representative sample. Interviews were conducted between April and August 2025 and used vignette-style prompts, example patient messages paired with draft responses that varied in tone, length, and implied authorship (AI versus human) to elicit how patients reasoned about AI involvement.
Several consistent findings emerged. Patients largely viewed portal messaging as transactional rather than relational a tool for quick answers, not deep interpersonal connection and this framing made many patients, including some who were otherwise wary of AI in medicine, more comfortable with AI drafting in this specific context than they were with AI in face-to-face care. Comfort with AI-drafted messages was described as generally high, but repeatedly conditioned on clinician review before sending; several participants said knowing a clinician had reviewed and edited the message "alleviates any fears." Preferences on tone and length varied widely and didn't split cleanly along whether a message read as AI-generated or human-written, some patients found longer, courteous AI-style phrasing reassuring, while others found it generic; one participant described an overly long draft by saying, "It's too long, just chill out. Nobody, no human, writes like that." Preferences also shifted with clinical stakes: patients wanted brief, efficient replies for low-stakes requests like prescription refills, but expected more detailed, supportive language or preferred a phone call altogether for higher-stakes messages, such as follow-up on a concerning test result. All participants said they wanted AI use disclosed, though they didn't converge on a single preferred wording or format for that disclosure, and several said disclosure language should avoid "marketing" phrasing and instead state plainly that AI was used and a clinician reviewed the message.
The study authors translated these findings into specific implementation recommendations for health systems, including: requiring human review before any AI-drafted message is sent, configuring EHR workflows so clinicians cannot send a message without reviewing it (with audit processes to monitor adherence), developing separate tone/length guidelines for low-stakes versus high-stakes message types rather than one universal template, and standardizing plain-language AI disclosure statements.
What This Means for Practices and Health Systems
For any practice or health system currently deploying or evaluating AI-drafted portal replies, this study offers a fairly specific checklist rather than a general endorsement or warning: build mandatory review into the workflow technically (not just as policy), don't rely on a single message template across all clinical contexts, and disclose AI involvement in plain, non-promotional language. The finding that patients don't want a single "one-size-fits-all" tone but do want a hard rule around clinician review suggests that governance effort may be better spent on the review/accountability workflow than on perfecting a single AI writing style.
What's Still Unsettled
This is a qualitative interview study with 40 participants drawn from a single academic health system. The authors themselves state it's not intended to produce statistically generalizable findings, and they found no clear, consistent patterns by age, race, or ethnicity in a sample this size.
It's also worth noting the study design: participants reacted to vignette-based example messages presented during a research interview, not to real AI-drafted messages they had personally received during actual care. How patients feel about a hypothetical message in an interview setting may not fully predict how they'd react to actual AI-drafted communication embedded in their own ongoing care. Finally, this is a follow-up to the same research team's earlier 2025 survey study, and one author (Dr. Eric Poon) disclosed a consulting relationship with a separate health-tech company (Triomics) outside this specific study disclosed by the authors themselves and not, from what's available, connected to the portal-messaging vendor or product being studied.
Sources:
Owens K, Jayaram A, Chowdhury A, et al. "Patient Perspectives on AI-Drafted Electronic Portal Messages." JAMA Network Open. 2026;9(7):e2622463. Published July 7, 2026. doi:10.1001/jamanetworkopen.2026.22463. https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2851278
This article summarizes a qualitative research study. It is not clinical or workflow-design advice; health systems should consult their own patient experience, informatics, and compliance teams before implementing AI-messaging policy changes.