This study employed a focus group methodology informed by co-design principles to explore patient and family member perspectives on the use of AI in clinical practice. Focus groups were selected to facilitate rich, interactive discussions, enabling participants to build on each other’s insights while generating diverse perspectives. While the primary aim was to gather feedback on specific case scenarios illustrating AI use in health care, the session was structured to go beyond simple elicitation of opinions. Co-design principles were incorporated to actively engage participants in collaboratively identifying key concerns, priorities, and desired safeguards for AI integration. Rather than developing a tangible product, the co-design focus centered on shaping participant-driven guidelines and recommendations for how AI should be implemented in ways that support patient-centered care.

A single 2-hour session was conducted in a centrally located research office with a large conference room and breakout rooms. The session combined full-group discussions with smaller, scenario-specific activities designed to promote collaboration and iterative refinement of ideas. This approach allowed participants to reflect on real-world examples while contributing to the development of contextually relevant strategies for responsible AI use in clinical settings.

Adult patients and family members aged 18‐80 years were recruited through email outreach, word of mouth, and networks such as our Patient and Family Advisory Council for Quality and Safety and the Georgetown University network. Recruitment focused on ensuring diversity in participant demographics, backgrounds, and health care experiences to capture a wide range of perspectives. Eligibility criteria included being English-speaking and willing to engage actively in the session.

The session was facilitated by a multidisciplinary team with expertise in human factors engineering, diagnostic safety, and patient engagement. Lead facilitators, trained in qualitative research methods, guided the session using a structured agenda designed to balance education, discussion, and cocreation. Recognizing the varying levels of participant familiarity with AI, the session began with an activity explicitly designed to surface and build a foundational understanding. This included group discussion of everyday AI examples (eg, navigation apps and virtual assistants) to ensure shared baseline knowledge before exploring health care–specific applications. We did not formally assess or quantify participants’ previous knowledge of AI or their digital literacy.

The session consisted of the following five key activities: (1) introduction and icebreaker (a brief overview of the session’s objectives, followed by an icebreaker to create a comfortable and engaging atmosphere), (2) “What is AI?” activity (participants discussed examples of AI in everyday life to build foundational understanding), (3) breakout sessions (participants were divided into smaller groups for 5 specific AI scenarios), (4) guideline or recommendation development, and (5) reflection and feedback (the session concluded with a debrief where participants shared insights and reflections). Materials provided to participants included an agenda, activity materials, and data collection for demographics and session evaluation.

In total, 5 scenarios were selected and developed to represent a diverse range of AI applications relevant to the diagnostic process and communication. These scenarios were designed to align with the study’s goal of examining patient perceptions of AI in diagnostic care by highlighting applications that varied in complexity, patient interaction, and clinical context. Each scenario was informed by a review of current AI use cases in health care and refined with input from our research team, including experts in diagnostic safety, human factors, and patient engagement. The five scenarios presented included: (1) portal messages (use of AI for patient portal messaging), (2) radiology review (use of AI in radiological imaging review), (3) digital scribe (ambient digital scribe for documentation in primary care), (4) virtual human (a virtual human presents a new diagnosis during a telehealth encounter), and (5) decision support (use of AI for clinical decision support to identify patients that would benefit from HIV testing) (Table 1). The scenarios were designed to elicit feedback based on different levels of patient interaction ranging from high interaction (virtual human and portal messages) where patients directly communicate with AI to indirect interaction (digital scribe) where AI is present during interaction with a human physician to minimal interaction (decision support and radiology review) where there is no direct communication between the AI and the patient. The specific scenarios can be found in Multimedia Appendix 1.

Breakout sessions included small group discussions (2‐4 participants) focused on the specific AI health care scenarios. Each scenario was presented by a dedicated facilitator who rotated between groups, ensuring that all participants discussed all 5 scenarios. Facilitators used standardized, prewritten scripts to introduce each scenario with a concise (approximately 1 minute) verbal description. To ensure consistent understanding, facilitators were prepared to clarify scenario details and answer participant questions as needed, using uniform prompts and clarifications.

For each scenario, facilitators guided the discussion using a structured set of questions designed to explore participants’ perspectives on that specific AI application in a clinical context. The first question asked participants to rate their comfort with the use of AI in the given scenario on a 1‐5 scale (1 being least comfortable and 5 being most comfortable). This question was explicitly framed to focus on the comfort level with the AI application as described in the scenario, not general attitudes toward AI. Additional questions probed what information participants would need to feel confident in the AI’s use, perceived benefits, potential concerns, and preferences for communication of AI-generated results. To promote consistency across discussions, facilitators received training on using the scripts, maintaining neutrality, and applying the structured question guide uniformly. Regular check-ins among facilitators during the session helped ensure alignment in approach and responses to participant questions. Facilitators also encouraged participants to share specific examples and personal experiences to enrich the discussion.

The session was recorded and transcribed verbatim, with facilitator notes collected to supplement the transcripts. A demographic questionnaire captured participant characteristics, including age, gender, health care experience, and self-reported medical conditions. Participants were asked to indicate any chronic illnesses or health conditions as part of a presession survey to better understand how their clinical experiences might inform their perspectives. In addition, participants were asked about their familiarity with AI and the frequency of AI use in their daily lives through structured survey questions, such as “Have you used AI applications like virtual assistants or automated systems? If so, how often?” These responses provided context for interpreting participant perspectives during the session. A postsession evaluation form gathered feedback on the session’s content, structure, and overall experience.

Thematic analysis was conducted using an inductive approach to identify patterns and themes within the focus group discussions. Furthermore, 2 researchers (GF and JB) independently coded the transcripts and resolved discrepancies through consensus. The initial coding process involved independently reviewing transcripts and assigning codes that captured key ideas and recurring sentiments. The coding team then iteratively refined and organized these codes into broader themes as a group. The process was informed by the discussion prompts, with themes often reflecting areas of interest, such as trust, communication, and perceived benefits or concerns. However, the themes were not strictly limited to the prompts, as additional insights emerged organically from participant discussions. To ensure the validity and relevance of the findings, the initial themes were presented to a patient-led steering committee as part of our AHRQ-funded Patient-Partnered Diagnostic Center of Excellence. This committee, comprising patient advocates and representatives, reviewed the themes, validated the findings, and provided additional feedback and considerations that were incorporated into the final analysis. Analytical memos documented the rationale for decisions and theme development throughout this iterative process. The final themes were organized to highlight both scenario-specific findings and cross-cutting issues, ensuring a comprehensive understanding of patient perspectives on AI applications in diagnostic communication.

This study received institutional review board approval by the MedStar Health Research Institute (STUDY00005888), and participation was voluntary. Informed consent was waived under the approved protocol. Participants were provided a US $100 gift card as compensation for their time and contributions during the 2-hour session. All data collected were deidentified prior to analysis to protect participant privacy and confidentiality. No personally identifiable information (PII) was retained or linked to study records. Data were securely stored on password-protected servers accessible only to the research team. These procedures were implemented to ensure compliance with ethical standards for human subjects research, including safeguards for confidentiality and privacy.

A total of 17 participants attended the AI focus group session, representing a diverse range of perspectives and varied experiences with health care (Table 2).

Participants reported a variety of medical conditions, reflecting a diverse range of health experiences. These included chronic conditions such as polycystic ovary syndrome, generalized anxiety and depression, hypertension, ulcerative colitis, arthritis, and diabetes. More complex conditions were also represented, such as avascular necrosis, stroke, kidney transplant, heart transplant, cancer, and post-traumatic stress disorder. This range of conditions provided valuable perspectives on the integration of AI in addressing diverse health care needs.

Participants expressed overall comfort with AI being integrated into the diagnostic process, as long as implementation involved key themes that addressed their concerns and expectations (Table 3, Figure 1). However, participants’ comfort levels varied significantly depending on the level of human interaction involved in the AI scenario. The results showed that comfort level drastically decreased as the amount of human interaction decreased in the AI process. For example, the scenario with which participants were least comfortable was the virtual human telehealth visit in which an AI-generated human would replace the physician when communicating a new diagnosis. Similarly, participants also appeared less comfortable with an AI chatbot sharing details about laboratory results. In contrast, participants were most comfortable with the ambient digital scribe scenario, in which an AI scribe documents clinical notes during a patient visit.

Across scenarios, participants expressed a mix of enthusiasm and caution, highlighting specific concerns around AI’s accuracy, transparency, and ability to meet individual patient needs.

Analysis of the co-design session discussions revealed 5 key themes that highlighted participants’ concerns, expectations, and opportunities for AI integration into clinical workflows: validation, usability, opportunities, transparency, and privacy (Figure 2). These themes provide critical insights into how patients perceive and evaluate AI technologies, which were further reflected in their comfort levels across different AI implementation scenarios.

Participants provided largely positive evaluations of the co-design session (Table 4).

Participants highlighted several aspects of the workshop that they liked best. They appreciated the open and nonjudgmental environment, which allowed for free sharing of opinions and thoughts without bias or pressure. Many valued the interactive nature of the session, particularly the small group discussions, which facilitated deeper engagement, diverse perspectives, and meaningful participation. The diverse backgrounds of participants, including patients from different races and professions, enriched the discussions and provided new insights. Participants also found the materials well-prepared, appreciated the brief AI introduction, and enjoyed the opportunity to learn more about AI in relation to their health care. Overall, the combination of open dialogue, group interaction, and thoughtful organization was highly praised.

Participants shared a few areas for improvement in the workshop. The most common concern was the limited time available, with several noting the need for more time to discuss topics in greater depth and brainstorm ideas. Some also suggested dedicating additional time for group discussions and addressing specific examples of AI currently in use or relevant case studies. Suggestions included incorporating more complex cases or scenarios, discussing AI bias in greater detail, and diversifying both the researcher backgrounds and participant groups to include more primary care providers, individuals from different socioeconomic groups, and a broader generational representation. While some recommended separating patients based on their AI knowledge for tailored discussions, others emphasized maintaining a mix of diverse perspectives within groups. Overall, participants highlighted opportunities to enhance inclusivity, depth of discussion, and time for meaningful engagement.

AI, while previously a technology of the future, has become a technology of the present. AI-driven technologies, including machine learning-driven decision support algorithms, deep-learning radiology scan classifiers, and large language model–driven digital scribes, have already been implemented in hundreds of hospitals nationwide. This study, in examining patient perceptions of 5 different scenarios describing current and future AI technologies in health care, provides a contemporary view of the multifaceted patient perspectives on AI’s role in providing diagnostic information, facilitating communication, and supporting decision-making. Many patient perceptions held true across all scenarios. First, a need for transparency in the development and validation of AI models, including their ability to reliably address the diverse needs of patients. Second, a preference for AI to complement rather than replace human providers, with an emphasis on maintaining human oversight in clinical decision-making. Third, the importance of clear and respectful communication about AI’s role in care, including obtaining patient consent, was seen as essential for building trust. Fourth, the potential for AI to enhance patient engagement, understanding, and access to information, provided it is implemented as a supportive tool that respects patient autonomy. Finally, concerns regarding the security and privacy of patient data highlight the need for transparency and robust safeguards to prevent misuse or unauthorized access.

The outlook on AI implementation into the diagnostic process was generally positive, with participants consistently highlighting AI’s ability to identify patterns and provide supplemental diagnostic information that might otherwise be overlooked by human providers. However, they emphasized that AI tools must integrate seamlessly into clinical workflows and preserve the essential human connection in patient-provider communication, as seen in the results regarding patient comfort levels across scenarios. Participants reported lower average comfort levels with high AI-patient interaction scenarios. For example, the scenario participants were least comfortable with was the virtual human scenario in which an AI-generated human would replace the physician when communicating a new diagnosis. Similarly, participants also appeared less comfortable with an AI chatbot sharing details about laboratory results. In contrast, participants were most comfortable with the digital scribe scenario, in which the application of AI was intended to enhance patient-provider communication by removing the need for providers to focus on documentation during clinic visits.

These findings highlight the importance of understanding patient perspectives within their specific health care contexts, offering insights into how AI can be integrated to enhance diagnostic processes and communication. For example, rural communities, where health care access and infrastructure often differ from urban settings, may present unique opportunities for AI to address gaps in care. Designing AI systems that are adaptable to varying levels of digital literacy, resource availability, and cultural expectations can help ensure these tools are both effective and equitable across diverse populations. Similarly, the interactive focus group discussions demonstrated the value of engaging patients with varying experiences and levels of familiarity with health care technologies. This diversity of input underscores the potential for AI to be developed in ways that resonate with patients from different backgrounds, fostering trust and satisfaction. By actively seeking and incorporating a wide range of perspectives, AI tools can be tailored to address the specific needs of different communities, ultimately supporting a more inclusive and patient-centered approach to care.

The findings of our work that held true across all scenarios are expectedly consistent with previous work – patients’ concerns with privacy, data security, and bias have been well-documented [26,27]. Specifically, our findings align well with the findings that patients have a generally positive outlook on AI’s implementation into their care as long as there are adequate guardrails to protect against a variety of potential harms [28]. Our work was unique in its focus on how such concerns are explicitly perceived in the context of AI applied to diagnosis and its communication in a variety of clinical contexts. Few qualitative studies have explored patient perceptions of AI in the context of diagnosis and communication, but our results appear to be in alignment with previous findings. Patients considered AI to be a helpful supplementary tool that should not serve as a replacement to human clinicians, a sentiment already documented for applications of AI in radiology [23].

While there is previous work identifying patient perceptions on the implementation of AI in health care broadly, there has been limited work identifying patient perceptions of AI’s role in reducing diagnostic errors through the enhancement of patient-provider communication. A recent scoping review identified that patients’ attitudes toward AI (which may impact their experiences when they interact with these tools in practice) are influenced by various factors, including familiarity with function, previous exposure to similar tools, supervision during use, and tool simplicity, validity, and cost. In light of this, it is imperative to consider patient perceptions of AI applications in a variety of clinical workflows in the context of diagnostic communication [17]. As AI becomes ubiquitous both inside and outside of health care, patients’ familiarity (and thus their attitudes toward AI) will continue to evolve. It is critical that patient perceptions of AI tools are continually assessed and used to enhance the diagnostic process and communication so that AI can be designed and integrated into the health care system in such a way that maximizes patient care and satisfaction.

Our findings have clear implications for clinical practice, finding that patients consistently stressed the importance of clinicians playing a central role in facilitating their experience with AI tools. Patients emphasized that their trust in AI would be built through transparent communication and clinician endorsement. Patients valued clinician involvement in contextualizing AI’s outputs, interpreting its recommendations, and providing assurances about its accuracy and reliability. Patients additionally expressed concerns that AI could disrupt workflows or reduce human interaction, particularly in scenarios where key diagnostic information was to be communicated. To address this, organizations should prioritize the implementation of AI tools that enhance human connection in care, such as those that reduce administrative burdens. These tools should be designed in ways that enhance, rather than overshadow, the clinician’s role in communicating diagnoses.

From a policy perspective, our findings reinforce the urgency of addressing gaps in regulations governing AI in health care, particularly concerning equity [29]. Patients expressed concern that biased data inputs could undermine the diagnostic accuracy of AI tools and have harmful effects on historically underserved populations. Policies should ensure that tools are trained on diverse datasets and are validated across representative populations in order to build patient trust and acceptance. Patients also expressed calls for transparency and informed consent regarding how health data is collected and used by AI systems in health care, aligning well with findings from a study [30]. These concerns should be addressed by policies that establish clear standards for disclosing when AI is used in care, and for ensuring that patients understand what personal information is being used, how it contributes to the diagnostic process, and how it influences the outputs from AI systems. Strengthening privacy and communication protocols will not only help address these concerns but also reinforce patient autonomy and trust in AI by supporting its ethical implementation in clinical settings.

Our findings underscore the critical role patients play in the acceptance and success of AI tools designed to enhance the diagnostic process, emphasizing the need to involve them in the development and implementation of these technologies. As primary stakeholders most directly impacted by changes to diagnostic workflows, patients have invaluable insights that can guide the design of tools to align with their expectations and foster trust [31]. Participants in this study expressed a dynamic view of AI tool implementation, with key insights into concerns that should be addressed during design and implementation, such as the importance of human connection and interaction, concerns regarding equity, personalization, and data security, and the pivotal role clinicians have in their understanding and comfort with new technology. By involving patients in the development of these tools, the health care system can better anticipate risks, communicate with patients more effectively, and deploy tools that not only improve the diagnostic process but also enhance trust and adoption, ensuring alignment with patient values and priorities.

This study has several limitations. First, focus groups were guided by different facilitators, which may have introduced variability in discussions due to differences in facilitation styles. The breakout sessions included only 2-4 participants per group, which allowed for in-depth discussions but may have limited the diversity of viewpoints within each session. This could affect the robustness of the findings; however, insights were aggregated across groups to capture broader themes. Future studies could address this by increasing group sizes or incorporating complementary methods such as individual interviews. Time constraints limited the discussion duration for each scenario, potentially restricting exploration of nuanced perspectives and the ability to achieve thematic saturation. In addition, the use of specific diagnostic-related AI scenarios provides structure but limits the generalizability of findings to other clinical contexts.

Participants were recruited through advisory networks, which may have introduced selection bias favoring individuals with an interest or familiarity in health care technologies. The participant demographics also reflect limitations in diversity, as more than 70% (n=13) were women, almost 65% (n=11) held graduate degrees, and none identified as having a Hispanic background. In addition, no participants reported high school or lower as their highest education level. According to national data, approximately 62% (n=115,011) of individuals aged 25 years and older in the United States have not attained a bachelor’s degree, suggesting that our sample overrepresented highly educated individuals [32]. Perspectives from participants with less formal education or from underrepresented backgrounds may have differed significantly, potentially revealing lower trust in AI, different concerns about its use, or alternate expectations for its role in health care. This lack of demographic diversity may limit the generalizability of the findings, as perspectives from individuals with different educational or cultural backgrounds could provide unique insights into AI applications in health care. We also did not assess baseline AI familiarity or digital literacy, which may have influenced participant engagement. Finally, this study is not comprehensive of all patient concerns about AI, with a focus on diagnostic applications shaping the discussions. Future research should aim to include a more demographically representative sample, explicitly assess AI literacy, and explore a wider range of clinical and nonclinical AI use cases to better understand how diverse patient populations perceive and respond to its implementation in the health care setting.

This study highlights the nuanced perspectives of patients on the use of AI in health care, with a particular focus on diagnostic communication. While participants recognized the potential of AI to improve diagnostic accuracy, efficiency, and equity, they also voiced significant concerns about transparency, trust, and the preservation of human connection. These findings underscore the importance of ensuring that AI tools are developed and integrated in ways that align with patient values and priorities. Key patient-oriented considerations include the need for clear communication about AI’s role in care, consent processes for its use, and opportunities for patients to actively participate in its development and implementation. Participants emphasized the importance of maintaining provider oversight, fostering understanding through accessible explanations, and designing systems that prioritize inclusivity and respect for patient autonomy.

As AI technologies continue to evolve and permeate health care, it is essential to iteratively assess and incorporate patient feedback to ensure these tools not only meet technical and clinical standards but also uphold the values of equity, transparency, and shared decision-making. By centering patients in the design and deployment of AI, we can create systems that not only enhance health care delivery but also foster trust and meaningful engagement between patients and providers.