Patient Perspectives on AI-Drafted Electronic Portal Messages

Key Points

Question

How do patients understand, interpret, and value artificial intelligence (AI)-drafted messages in online patient portals?

Findings

In this qualitative study of 40 patients from a large academic health system, patients valued efficiency, strongly preferred clinician review of AI-drafted messages, and wanted disclosure of AI use to maintain trust. Patient preferences regarding ideal AI tone and message content varied widely, and most participants perceived digital messaging as transactional rather than an integral part of the patient-clinician relationship.

Meaning

These findings suggest that AI-assisted messaging systems should prioritize efficiency, context sensitivity, clinician oversight, and clear disclosure that reflects patient communication goals.

This qualitative study examines patient perspectives about messages drafted by artificial intelligence (AI) in online patient portals to identify implications for patient-centered implementation.

Abstract

Importance

Patient portal messaging volume has increased substantially and is linked to clinician burnout. In response, health systems are increasingly adopting large language models (LLMs) to generate draft replies to patient messages, yet little is known about how patients interpret and evaluate artificial intelligence (AI) involvement in this communication channel.

Objective

To understand how patients perceive AI-drafted responses to electronic patient portal messages and to identify implications for patient-centered implementation.

Design, Setting, and Participants

This qualitative study used vignette-based prompts mirroring scenarios from a prior survey on patient preferences regarding AI-drafted online messaging. Participants included a sample of 40 adult patients from a large academic health system who had previously been surveyed about AI-drafted portal messaging. Recruitment sought increased representation of individuals from minoritized racial and ethnic groups, older patients, and participants with lower satisfaction with AI-drafted messaging in prior survey responses. Data were collected between April and August 2025.

Exposure

One-time interviews conducted via videoconference. Interviews included vignettes and draft messages varying in tone, length, and implied authorship, alongside questions about messaging preferences, AI, and health care.

Main Outcomes and Measures

Patient perspectives regarding (1) portal messaging preferences, (2) comfort with AI-drafted replies, (3) preferred message tone and content, and (4) AI disclosure expectations.

Results

This study of 40 patients included 30 women (75.0%), 9 men (22.5%), 14 Black or African American participants (35.0%), 13 White participants (32.5%), and 13 patients aged 65 years or older (32.5%). Patients described portal messaging as transactional, prioritizing timely problem resolution over relational depth. Patients expressed high comfort with AI-drafted messages, conditional on clinician oversight. Patient preferences for tone and expressions of empathy varied across individuals and contexts and were driven less by whether messages appeared AI-like or human-like than by whether tone, length, and detail fit the purpose and stakes of the communication. Participants broadly endorsed AI disclosure, describing transparency as important for trust, but differed in preferred disclosure timing and format.

Conclusions and Relevance

In this qualitative study of patient perspectives on AI-drafted portal messaging, patients supported AI use when it improved communication efficiency, so long as clinicians reviewed and remained accountable for messages. Implementation should prioritize clinician oversight, context-sensitive communication standards, standardized disclosure practices, and monitoring for downstream effects on quality, equity, and patient trust and safety.

Introduction

Electronic patient portals have become routine infrastructure for outpatient care, enabling patients to ask clinical questions, request refills, and manage ongoing care between visits.¹ At the same time, the rapid rise in portal-based communication has produced a growing volume of clinician in-basket messages that substantially contributes to administrative workload and is linked to clinician burnout and dissatisfaction.^2,3 Health systems seek scalable strategies to manage this workload while preserving timely communication and care quality. As such, many are turning to generative artificial intelligence (AI) and large language models (LLMs) to draft proposed responses to patient messages, which can then be reviewed, edited, and sent by clinicians.⁴ These tools are being integrated directly into electronic health records and clinical workflows, positioning AI as a key part of everyday patient-facing communication.⁵

The quick uptake of LLM-generated drafts into patient messaging portals raises questions that extend beyond efficiency.⁶ Patient portal messages are sites of information exchange, clinical guidance, and interpersonal care, and patients may interpret message tone, length, and responsiveness as meaningful indicators of empathy, attentiveness, and professionalism.^7,8,9 In our prior survey study examining patient preferences for AI-mediated portal messaging systems, participants generally preferred AI-drafted responses over clinician-written responses and rated them as more satisfactory across multiple message scenarios.¹⁰ At the same time, satisfaction and perceived care were modestly lower when participants were explicitly informed that a response was authored by AI. This paradox—preference for AI-drafted messages alongside reduced ratings when AI involvement is disclosed—suggests that patient evaluations are not only shaped by the content of messages but by beliefs about what AI use means in the context of clinical relationships and the delivery of health care.

While our prior research identified broad patterns of patient preference and reactions to AI disclosure, it could not explain the reasoning underlying these judgments. As a result, key ethical and operational questions remain unresolved, including how patients conceptualize acceptable use of AI in patient-facing communication, how they weigh efficiency and responsiveness against human relational expectations, what norms of disclosure and transparency feel appropriate, and how AI-mediated messaging may affect patient-clinician relationships or the trustworthiness of health care institutions. These unanswered questions are particularly salient as health systems move from experimentation to routine implementation of AI drafting tools, often with limited guidance or patient input. Accordingly, the objective of this qualitative study was to understand how patients make sense of AI-drafted online messaging technologies in patient portals and to identify implications for the patient-centered implementation of these tools in health system communication workflows.

Methods

Study Design and Ethics

We conducted a qualitative study using in-depth, semistructured interviews to examine how patients understand the use of generative AI to draft replies to patient portal messages. This was a multi-institutional, collaborative study designed as a qualitative follow-up to our previously published survey on patient preferences regarding AI-drafted portal messaging and disclosure practices.¹⁰ Accordingly, our sampling frame and interview vignettes were directly derived from the survey instrument and clinical scenarios. Our study design and methods follow the Consolidated Criteria for Reporting Qualitative Research (COREQ) reporting guideline.¹¹ Data were collected between April and August 2025.

This qualitative study was determined to be exempt from review by the Duke University Health System Institutional Review Board and the New York University Grossman School of Medicine Institutional Review Board. Potential participants received written information about this study via email and an online Qualtrics form during the recruitment process. Before enrolling in the study, participants completed a verbal consent process that included permission to record and publish anonymized participant quotes. All participants received $50 gift cards.

Participants and Recruitment

Participants were adult patients from a large academic health system who had previously completed a patient advisory committee survey regarding patient portal messaging and AI-drafted responses. We used purposive sampling to recruit a diverse group of respondents with variation in demographic characteristics and survey responses. Recruitment proceeded in 2 phases. First, we contacted surveyed participants representing a wide range of views, with oversampling to increase representation of patients from racially or ethnically minoritized groups and older patients, whom we hypothesized might hold distinct expectations about trust, communication, and technology in health care. Although we did not purposively oversample by gender, women were more likely to agree to participate. Second, we conducted an additional targeted recruitment effort focused specifically on participants who expressed lower satisfaction with AI-drafted patient messages in the prior survey, with the goal of strengthening viewpoint diversity and better understanding critical or skeptical perspectives.

Data Collection

We conducted interviews with individual participants via videoconferencing (Zoom) that lasted approximately 45 to 60 minutes. We used a semistructured guide (eAppendix in Supplement 1) and incorporated vignette-based prompts that mirrored key scenarios from the prior survey study (eg, routine medication refill requests, prescription side effects, potentially abnormal test results). After initial open-ended questions, participants were presented with standardized vignette scenarios consisting of example patient messages and paired draft responses that varied in tone, length, and implied authorship (AI vs human), which were discussed iteratively to elicit participants’ interpretations, preferences, and reasoning about AI involvement and disclosure. We audio recorded and transcribed all interviews.

Data Analysis

We analyzed interview data using an abductive analytic approach, a framework popular in medical sociology that iteratively develops and refines themes through movement between inductive insights from the data and sensitizing concepts from existing scholarship.¹² A multidisciplinary team of 4 coders conducted analysis. The team developed an initial codebook based on the interview guide and early collective review of 4 transcripts, then iteratively refined codes through team discussion and repeated engagement with transcripts. To support analytic rigor and consistency, at least 2 coders reviewed each transcript, and the team met to resolve disagreement through consensus. We developed themes through iterative memo-writing, comparison across participants, and attention to patterns in how patients reasoned about AI involvement across scenarios. We used Dedoose 10.0.25 (Dedoose) qualitative analysis software for data management, coding, and retrieval of coded excerpts, but theme development was not automated.

Results

This study of 40 participants included 30 women (75.0%) and 9 men (22.5%); 14 participants (35.0%) identified as Black or African American, 13 (32.5%) as White, and 13 (32.5%) were aged 65 years or older (Table 1). Our findings include the major themes from patient interviews regarding perceptions of AI-drafted responses to portal messages. We first describe patient attitudes toward portal messaging as a communication modality, followed by findings on comfort with AI-drafted messages, preferences regarding message tone and content, and perspectives on disclosure and transparency. Representative quotations illustrating each theme are provided in Table 2.

Table 1. Participant Demographics.

Characteristic	No. (%), (N = 40)
Gender identity
Male	9 (22.5)
Female	30 (75.0)
Unknown	1 (2.5)
Race or ethnicity
American Indian or Alaska Native	0
Asian	3 (7.5)
Black or African American	14 (35.0)
Native Hawaiian or Pacific Islander	0
White	13 (32.5)
Hispanic	3 (7.5)
More than 1 race	2 (5.0)
Other	3 (7.5)
Unknown	2 (5.0)
Age group, y
18-24	0
25-34	3 (7.5)
35-44	5 (12.5)
45-54	7 (17.5)
55-64	10 (25.0)
65-74	8 (20.0)
75-84	5 (12.5)
≥85	0
Unknown	2 (5.0)

Table 2. Themes, Representative Quotations, and Analysis Regarding Patient Attitudes Toward AI-Drafted Portal Messages.

Patient perspectives	Representative quotations
Attitudes toward online portal messaging
Messages were perceived as transactional, not relational	“[I’m looking for] A resolution to a situation.”
	“I’m looking for quick information…I don’t have time for this long thing.”
	“It doesn’t really need to be very personal. It just needs to be quick and to the point...You see that I’m having issues, and how do I address them?”
	“I don’t have time for long emails. And when you go on your online portal you know I’m not going there to be loved...I’m going on there for you to help take care of me, you know...I just want the answer. I want you to take care of me and what’s going on.”
	“The messaging app is useful for brief questions, quick responses, things like that. I don’t expect to solve big issues through MyChart and Epic, I expect to have simple exchanges there.”
Patients less comfortable with AI use in medicine did not mind the use of AI in online messaging	“As a patient, I’m more nervous about [AI]. I think for the doctors it’s probably good for them, probably saves time for them...I’m comfortable for the doctor, but not for myself particularly. I know it saves time for them, but I think we’re losing a little patient contact, [it’s a] little too automatic, too robotic. One of [my] doctors...he came in with his phone and he said, ‘Do you mind if I record this?’ So he did, and then he showed how AI had put it all into the appointment summary...What’s gonna happen to that recording, you know? And because you are being recorded, it makes you nervous…I think [it] cuts down on the personal touch.”
	When the same patient was asked about the “personal touch” regarding AI-drafted online messages: “I’m only looking for [a personal touch] in person, not online...that’s online...it’s a machine, so I don’t expect the interpersonal.”
Comfort with AI-drafted messaging
Comfort was generally high	“I would feel comfortable [with AI-drafted messages].”
	“I know doctors are busy…Even though responding to patients is very important, doctors also have a lot of other urgent issues and priorities that they have to handle. AI generated responses for the doctors will…save them time and allow them to be able to respond to patients quickly while being able to also handle other priorities during the day…I feel like if AI can generate a message that the doctors feel comfortable with, and it answers all the questions that the patients have, it’s a win for both the doctors and for the patients, because we’re going to get responded to quicker because they’re not having to sit there and generate a message from scratch and AI is going to help them to keep things moving and flowing quicker.”
Conditional on clinician oversight	“I would be okay if…the reply, the message…started off by being AI generated, and then it was reviewed by the primary care provider and sent along to me…The fact that artificial intelligence originally drafted the reply, 1st of all: would not surprise me, and Number 2: would not alarm me in any way.”
	“I think as long as there’s an indication…that somebody reviewed it before they sent it, that alleviates any fears.”
	“Knowing that a doctor has reviewed [the message] and edited it would be helpful.”
Message tone and content
Tone
Many patients thought the AI-drafted messages were too long, generic, or impersonal.	“I think [the AI-drafted message] seems a little bit lengthy. I’m more of a concise person, so I think the first paragraph or sentences could be reduced just to simplify. The first part makes it feel a little auto-generated and not as personable, but overall, the rest of it is good.”
	“It’s too long, just chill out. Nobody, no human, writes like that.”
Others appreciated the courteous, customer service tone of AI-drafted messages.	“I can see that this is like a form message. I really like the form of it...I really liked kindness and leading with kindness...and I feel like this message is very kind and like affirming, and also invites further conversation if you have additional questions, and like its just like polite customer service kind of scripts that is comforting.”
	“I like ‘We appreciate your diligence in managing your medication.’ I think that’s important...more friendly.”
Empathy
Empathy was not equated with message length or niceness; empathy emerged from responsiveness and situational fit, not affective language.	“I think the brevity of the second one makes it feel a little more empathetic just because it doesn’t feel like it’s just copied and pasted from a website…it feels a little more like somebody sat down and wrote the message.”
Others found the affective language from AI-drafted messages empathetic, even when they knew the message was drafted by AI	“First, I would say...[the AI-drafted message] shows empathy because...it’s first of all apologizing which makes me feel like they care that I’m having an issue...And then it’s very detailed...Then the final 2 sentences, ‘Your health and well-being are important to us, and we appreciate your diligence in managing your medication...’ It makes me feel like the hospital or the doctor cares because it’s saying that my health and well-being are important, and it’s wishing me continued good health. So I feel like this message, it makes me feel cared about. It makes me feel like this doctor or this office is really concerned and really invested in my well-being. Right? So after reading this, I would feel pretty happy about this message.”
	“I would much rather receive this very caring, detailed, thought-out AI-generated message opposed to [the human response] that is cut and dry and reads like: ‘I don’t have time…I don’t really know your specific health situation, I’m just sending you a message to get it out there…and I can move on to something else.’ So I still feel very comfortable and very warm about this response, even if it was AI-generated. Now, if my doctor did not have to read it and was not allowed to make changes, and did not have to approve it before it went out, then my feelings about it would be different. But the doctor is reading, is approving, is making changes…So the fact that it’s AI-generated does not change my feelings about it.”
Patients did not prioritize empathy in online messaging	“I usually just want to get an answer. I don’t need all of the extra stuff. I’m not expecting (empathy) that I might expect if I were in the office with my doctor directly, but from just messaging…I don’t need the empathy. I don’t need ‘I’m sorry that you’re feeling this way’ or anything like that. I just want the answers at that point.”
Context
Short messaging was preferred for lower stakes interactions (eg, prescription refills); longer messaging with detailed responses and affective language was preferred for higher stakes interactions, (eg, interpreting ambiguous test results).	Two responses from same participant: (Response 1) “[The AI-drafted response for a prescription refill] a really nice response. It’s just way too long…All I want to know is, did you see my message? Did you react to it? But all of this apology, I don’t need all of that…If it was something pertaining to my health, then maybe saying that would be comforting. But for asking for a prescription, I don’t need all of that…but if they’re re-looking over my mammogram, and they saw something and want me to come back now…that’s important to me.” (Response 2) [Re: an AI-drafted message for prescription refills] “I find it profoundly annoying, because it’s all this gobbledygook words that I know don’t come from a real person…it’s inauthentic. It’s completely the wrong path to go down.”
	[Re: an AI-drafted message about concerning laboratory results] “Here, where stakes are high, this patient is clearly…highly worried or somewhat anxious, because there was a [concerning laboratory report]. So I would hope that every doctor would be a little more careful in an answer...right? So this deserves, I’m more accepting of AI helping here, when a really busy doctor might really muff it, not on purpose, just because they don’t have time to write the kind words.”
Other forms of communication were preferred for serious medical questions	[Re: vignette on laboratory report interpretation] “This one would be a call for me…I would call and see if I could have 5 min to talk to somebody to determine if I need to go in or not…By having someone that I can talk with directly, now we can have a conversation back and forth vs: I send a message, I wait, they respond, I see it, finally, I respond, I wait. Whereas, [on the phone] we can go back and forth with the communication, and I can have all my questions answered upfront…So it’d be good if we could just have a little face time…Again, I would rather be speaking with someone.”
Disclosure and transparency
Patients wanted disclosure of AI use, and disclosure preferences varied	“I would like to be informed if the information was AI generated.”
	Interviewer: “Would you want to know [AI] is being used [in portal messaging]?” Patient: “Every single time, for every patient, without any exception, ever.”
	“None of the [disclosure messages] really stood out.”
	“I need to be reminded. So definitely I’d like to be told…from the front desk or from whoever when I’m checking out that [messages] are going to be AI generated. And then, when the [message] comes, they could put that little [disclosure statement] at the bottom as well.”

Abbreviation: AI, artificial intelligence.

Patient Attitudes Toward Online Portal Messaging Systems

Across interviews, participants described online portal messaging as a useful modality for managing routine clinical needs but not as a setting where they expected deep interpersonal connection. Most participants framed portal messaging as fundamentally transactional—an efficient mechanism for obtaining information, resolving questions, or completing administrative tasks. Patients emphasized that the primary goal of portal messaging was “a resolution to a situation” and “quick information,” often describing concise replies as preferable and sufficient for meeting their needs. Consistent with this framing, participants reported that they did not expect portal messages to be personal and that they generally viewed the portal as a space for brief exchanges rather than complex or emotionally intensive discussions.

Notably, even participants who expressed broader unease about AI in medicine emphasized that personal touch mattered most in face-to-face encounters, whereas online messaging already felt transactional and impersonal. As a result, these AI-hesitant participants viewed AI involvement in portal messages as more acceptable than AI use in direct clinical decision-making or in-person interactions, such as ambient documentation systems.

Comfort with AI-Drafted Online Messages

Participants generally expressed comfort with AI-drafted portal messages and frequently perceived these tools as a reasonable response to clinician workload. Many patients recognized clinicians as “busy” and interpreted AI-generated drafts as a means to help clinicians respond more quickly while managing competing priorities. In this framing, AI-supported messaging was viewed as mutually beneficial because it could reduce clinicians’ burden while increasing the timeliness of patient responses.

However, participants’ comfort was consistently described as conditional rather than absolute. Patients repeatedly emphasized that AI drafting was acceptable so long as messages were reviewed and approved by a clinician before being sent, even for lower-stakes tasks. Several participants described clinician review as the key safeguard that “alleviates fears,” reinforcing that patients viewed portal messaging as an extension of clinical responsibility even when messages were treated as transactional.

Patient Perspectives on Message Tone and Content

Participants offered heterogeneous perspectives on what constituted an acceptable portal message. Preferences varied widely and were shaped by the clinical context and perceived stakes of the message. Many participants evaluated message tone and content through practical criteria, including clarity, concision, relevance to the question, and whether the reply appeared to be responsive rather than generic. In particular, some patients described AI-generated drafts as overly long or impersonal, with language that felt “auto-generated,” “form-like,” or overly scripted. For these participants, complex responses could be interpreted as inefficient, unnecessary, or even inauthentic, with one participant remarking, “It’s too long, just chill out. Nobody, no human, writes like that.”

In contrast, other participants valued features often associated with AI drafts, including a courteous, reassuring, “customer service” tone. Several appreciated the politeness and affirming language, such as acknowledging patient effort (“we appreciate your diligence in managing your medication”) and inviting follow-up questions (“please don’t hesitate to contact us”). For these patients, supportive phrasing served as an indicator of attentiveness and care, and message length was sometimes interpreted positively when it conveyed thoroughness and clinical guidance.

While empathic online communication was not a primary goal articulated, participants also did not converge on a single shared definition of perceived empathy in portal communication. Many patients did not equate empathy with message length, formality, or explicitly affective language. Instead, participants discussed empathy as responsiveness and situational fit—whether the message appropriately matched the patient’s concern and level of urgency. Some participants explicitly preferred brief, direct replies, emphasizing that the purpose of messaging was to obtain answers rather than emotional support. For these participants, stereotypically empathic language, such as “your health and well-being are important to us” could feel unnecessary or mismatched to the task, particularly in low-stakes situations, such as prescription refills. Others described brevity itself as a signal of authenticity and care, suggesting that shorter replies felt less like copy-paste text and more like something a person “sat down and wrote.” Several participants also noted that empathic language was not diminished by AI involvement, describing AI as an efficiency tool where, as long as clinicians reviewed and endorsed the message, the resulting expressions of care were understood as coming from the clinician.

Participants also stressed that their preferences on message tone and content depended on clinical stakes. For routine tasks (eg, prescription refills), overly nice or apologetic language could feel excessive or irritating. However, for messages involving uncertainty, anxiety, or potentially concerning test results, many participants expected clinicians to communicate with greater care and reassurance. For some participants, this meant that they would prefer clinicians to draft higher stakes messages or would prefer to use in-person communication methods. At the same time, some patients noted that AI assistance might be more acceptable in higher-stakes situations because it could help busy clinicians craft a thoughtful message. For example, a participant noted that an AI-drafted message felt “profoundly annoying” in a prescription refill scenario because it appeared inauthentic but described greater acceptance of AI language for concerning results when the patient was “highly worried,” reasoning that the situation deserves a careful and supportive response that a clinician may not have time to craft.

Generally, our findings indicate that patient responses to message tone and content were not primarily determined by whether a message was perceived as AI-like or human-like, but by whether tone, length, and detail aligned with the purpose and stakes of the communication. Participants’ preferences varied both across patients and within the same patient depending on different clinical scenarios.

Patient Perspectives on AI Disclosure and Transparency

Disclosure of AI use emerged as an area of broad consensus across participants, as all patients expressed a desire for transparency about AI use across health care contexts, not only in portal messaging, and generally did not report increased skepticism or altered interpretations of messages when AI involvement was disclosed. Rather, many described disclosure as fostering trust by signaling honesty and reducing concerns that AI use might otherwise feel hidden.

At the same time, when we showed participants a list of potential AI disclosure messages that could be included in patient messaging portals, most patients did not have strong preferences between them, instead reporting that one disclosure statements did not feel meaningfully different than any other. Importantly, some participants emphasized that disclosure should be clear and free of “marketing” jargon, such as “to improve your care experience,” favoring plain statements that AI was used and that a clinician reviewed and/or edited the message. Preferences varied regarding the timing and format of disclosure. Some participants preferred disclosure at multiple points in the care process, including notification during clinic check-out or via front desk staff, coupled with reminders appended to portal messages.

Discussion

Our study suggests that patients are generally comfortable with the use of generative AI to draft in-basket responses, particularly when framed as a workflow support that could improve response speed and reduce clinician burden. Patients emphasized that the primary value of portal communication is timely resolution of health-related questions rather than personalization or relational depth. This framing is imperative for health systems implementing AI-assisted messaging because it suggests that patients may evaluate these tools less in terms of whether AI replaces human connection between clinicians and patients and more in terms of whether AI improves reliability, clarity, and responsiveness.

Participants also described how expectations for message tone and content shifted with clinical context, with several expressing a preference for more detailed or emotionally supportive language in higher-stakes scenarios, such as interpretation of potentially concerning test results. In comparison, our prior survey results found that topic seriousness did not significantly alter patient preferences. This divergence may reflect methodological differences—surveys capture aggregate preferences across standardized scenarios, and qualitative interviews surface context-sensitive reasoning and situational expectations. Still, additional mixed-methods and studies in the clinical setting would help further understand patient preferences. Participants also broadly endorsed transparency about AI use, while varying in where, when, and how disclosure should occur.

Another central and consistent finding was that patient acceptance was conditional on clinician involvement. Participants repeatedly described AI drafting as acceptable so long as a clinician reviews, corrects, and remains accountable for the message. This aligns with a broader body of literature showing that patients tend to support AI as an adjunct to, rather than a replacement for, clinician judgment and that they expect clinicians to retain ultimate responsibility for care decisions and communications.^{13,14,15,16,17} We suggest that health care organizations should treat AI-assisted portal communication not as mere clerical automation but as a clinical communication intervention requiring explicit oversight norms about when edits are required and how escalation should occur when the clinical stakes are high.

Patients also expressed varied preferences regarding tone and message length, with disagreement about whether empathic language felt caring, inauthentic, or simply inefficient. The perceived empathy of online communication was often understood as responsiveness and situational fit, including brevity when appropriate, rather than affective language. These findings both align with and complicate a growing body of literature suggesting that AI-generated responses are frequently rated as more empathetic or caring than those produced by clinicians or other humans, including our prior survey.^{10,18,19,20,21} At the same time, this literature highlights a paradox in which individuals may rate AI-generated messages as more caring or empathetic while still preferring to receive empathy from human sources.^10,22,23 Our findings help reconcile this tension by suggesting that patients interpret and value empathy in context-dependent ways, often prioritizing whether communication meets their needs in a given situation over whether it reflects human-authored emotional expression. Many participants viewed AI-generated messages as capable of conveying care, but only when embedded within a system of clinician oversight and accountability. This finding supports the development of context-sensitive implementation guidelines rather than 1 ideal standard message, including differentiated approaches for lower-stakes transactional tasks (eg, prescription refills) vs higher-stakes or uncertainty-laden messaging (eg, interpreting imaging results) where longer, more supportive language may be welcomed. More broadly, our results suggest a need for future research to more deeply conceptualize what constitutes empathetic AI-mediated communication in health care and to examine how different forms of communication contribute to patients’ sense of being cared for across clinical contexts.

These findings suggest several policy and strategy implications for AI-drafted online messaging systems (Table 3). The ethical justification for AI-drafted portal messaging depends on whether this technology advances patient-centered goals like faster, clearer communication, while also improving clinician well-being and maintaining safeguards for safety and accountability. Accordingly, health systems implementing LLM-assisted drafting should treat these tools as clinician-facing workflow supports that continue to require clinician review prior to sending, with explicit institutional accountability for content accuracy and appropriateness. Our results further support the need for context-sensitive implementation guidelines informed by sustained patient engagement, including standards for tone, length, and expressions of empathy that may vary by message type, rather than a single communication template applied universally across clinical scenarios. In parallel, transparency should be standardized as an implementation norm, with clear, consistent disclosure language that indicates AI involvement and clinician review while avoiding marketing-like framing. Finally, because patient portal messaging is increasingly central to clinical care, implementation should be paired with ongoing monitoring of downstream outcomes.

Table 3. Translating Patient Perspectives Into Actionable Implementation Strategies for Artificial Intelligence (AI)-Drafted Portal Messaging.

Study finding	Action steps for health systems
Patients support AI-drafted messaging when it improves timeliness and efficiency without compromising care quality or safety	Define clear performance goals for AI drafting (eg, reduced response time, clinician workload)
	Regularly evaluate whether these tools meet those targets without introducing safety concerns
	Incorporate both patient-centered and clinician well-being metrics when assessing impact
Acceptance of AI drafting is conditional on clinician review and accountability	Require human review prior to sending all AI-drafted messages
	Establish clear responsibility for message content within care teams
	Configure electronic health record workflows to prevent clinicians from sending messages without review and develop audit processes to monitor adherence
Preferences for message tone, length, and empathy vary by clinical context and patient expectations	Develop context-sensitive communication guidelines that differentiate between low-stakes administrative messages and higher-stakes clinical communication
	Engage patients through Community Advisory Boards (CAB) or user testing to inform message templates, tone options, and drafting controls
	Allow flexibility in AI outputs rather than enforcing a single standardized style
Patients broadly support transparency about AI use but vary in preferred timing and format of disclosure	Establish standardized, plain-language disclosure practices indicating AI involvement and clinician review
	Implement a standard process for where disclosure occurs (eg, portal messages, clinic intake, institutional notices) and ensure consistent implementation across departments
	Periodically reassess disclosure language with patient input
Additional research is needed to assess how AI-assisted messaging may affect patient trust and communication quality over time	Implement ongoing monitoring of downstream outcomes, including patient satisfaction, perceived responsiveness, miscommunication events, and disparities in message quality
	Incorporate feedback loops using patient surveys, safety reporting systems, and equity audits to guide iterative improvement

Abbreviation: AI, artificial intelligence.

Our data suggests that patients evaluate portal messages for signals of authenticity and care based on their conscious or unconscious expectations for how these concepts are expressed in digital communication. Additional research would be helpful to understand the implicit values and criteria patients use to define empathy and care in health communication. Future research should also examine implementations in the clinical setting to assess how AI drafting affects communication quality, clinician oversight, equity, safety, and trust over time. For example, future work could explore what organizational, technical, or workflow safeguards effectively support clinician review and meaningful editing of AI-drafted messages, and how health systems can design processes that promote active oversight rather than passive forwarding of automated content.

Limitations

This study has several limitations. Because qualitative research is designed to provide in-depth, contextually grounded insights rather than to produce statistically generalizable findings, we do not expect that our relatively small sample drawn from a single health system will necessarily be applicable to all patient populations or settings. Although purposive sampling was used to increase viewpoint diversity, we could not discern any clear patterns across subgroups (eg, by age, race, or ethnicity), and larger studies will be needed to better characterize how preferences for AI-assisted messaging and clinician oversight vary by demographics or other characteristics such as education level, health status, digital literacy, or prior experiences with the health system.

Conclusions

In this qualitative study of patient perspectives on AI-drafted portal messaging, patients generally viewed AI-drafted portal messaging as acceptable and potentially beneficial when it supported timely, efficient responses to their questions. However, acceptance was conditional: participants emphasized that clinicians must remain accountable through review and oversight and that implementation should reflect the context-specific nature of digital communication, including variation in preferences for tone, length, and expressions of empathy across message types. Patients also broadly endorsed transparency about AI involvement. Together, these findings suggest that AI-assisted messaging should be implemented as a clinician-facing support tool rather than a substitute for clinical responsibility, with governance approaches that prioritize safety, context sensitivity, and patient engagement.

Supplement 1.

eAppendix. Interview Guide

Supplement 2.

Data Sharing Statement