An Exploration of the Promises and Perils of Responsible Deployment of Health AI for Safety Net Populations as Perceived by Healthcare Providers

An Exploration of the Promises and Perils of Responsible Deployment of Health AI for Safety Net Populations as Perceived by Healthcare Providers | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article An Exploration of the Promises and Perils of Responsible Deployment of Health AI for Safety Net Populations as Perceived by Healthcare Providers Ishani Purohit, Matt Kammer-Kerwick, Emily Spandikow, Gregory Pogue, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7013847/v2 This work is licensed under a CC BY 4.0 License Status: Posted Version 2 posted You are reading this latest preprint version Show more versions Abstract The current study examines the responsible deployment of AI in healthcare settings with a particular focus on underserved, safety net populations. We employ a mixed methods approach to study the perceptions of health care providers relative to the promises of responsible deployment of AI and the potential perils that need to be navigated to achieve that goal. As health AI tools continue to enter clinical spaces, understanding how they are perceived by providers in safety-net environments is essential for equitable implementation. This study revealed that while there is cautious optimism among healthcare professionals, particularly regarding improvements in workflow, personalization, and efficiency, significant concerns remain around data integrity, trust, and infrastructural readiness. High-trust providers viewed AI as a valuable support system, whereas low-trust providers raised critical questions about governance, privacy, and the risk of exacerbating existing inequities. The findings emphasize the importance of human-in-the-loop models, localized implementation strategies, and community-informed design to ensure that the promise of AI does not bypass the populations it seeks to serve. Moving forward, engaging providers in policy design, tool development, and implementation processes will be crucial for realizing the equitable integration of AI in healthcare. Scientific community and society/Business and industry Health sciences/Health care Scientific community and society/Scientific community Responsible AI Health Equity Provider Perceptions Safety Net Populations Mixed Methods Vignette Analysis Figures Figure 1 Introduction The current study examines the responsible deployment of AI in healthcare settings with a particular focus on underserved, safety net populations. These populations can be rural or urban and are typically characterized by their lack of access to healthcare services, lower-income and resource status, uninsured or under-insured status, and high prevalence of chronic disease. We employ a mixed methods approach to study the perceptions of health care providers relative to the promises of responsible deployment of AI and the potential perils that need to be navigated to achieve that goal. Specifically, we employed a two-study design that involved a set of exploratory interviews among thought leaders and practitioners followed by a survey of practitioners to broaden the perspectives included in our analysis. While the responsible deployment of AI in healthcare settings is important overall, we chose to focus on safety net populations because of some of the unique challenges faced by healthcare providers in serving those populations, including for example, high patient to provider ratios, significant constraints on resources, and the presence of greater health disparities among those community members than in the general population. The survey used in this exploratory study included both quantitative and qualitative questioning; our focus in this paper is on the qualitative insights obtained. It is important to note that the survey design included a series of structured vignettes that presented health care providers with specific realistic but fictional health AI technologies. Within these vignettes, healthcare practitioners provided ratings for the degree of appeal and the degree of concern associated with deploying the described technology among safety net community members. The survey also asked healthcare providers to provide a qualitative explanation of how they were thinking about responsible deployment of that technology in the context of those perceived points of appeal and points of concern. We performed a literature review that allowed us to focus on specific gaps and used the interviews to develop a more structured line of questioning for the survey. From this first phase we established the following research questions: RQ1: What benefits are currently perceived to health AI? RQ2: What concerns are currently perceived for health AI? RQ3: How are healthcare providers thinking about navigating these concerns in ways that allow the benefits of health AI to be achieved responsibly? We believe that responses to these questions from the perspectives of safety-net healthcare providers will provide unique insights regarding the design and deployment of health AI in general and in under-resourced settings, specifically. Literature Review Artificial intelligence has the potential to transform healthcare delivery, particularly in rural and underserved settings where resource limitations and workforce shortages create significant barriers to care. These settings, often characterized by high patient-to-provider ratios, fragmented healthcare infrastructure, and limited access to specialty services, stand to benefit greatly from AI-driven solutions aimed at improving diagnostic accuracy, clinical workflows, and patient management ( 1 , 2 ). However, the integration of AI into these environments presents unique technical, financial, and ethical challenges that must be addressed to ensure equitable access to high-quality care. Rural safety-net hospitals serve populations that are disproportionately low-income, uninsured, and burdened with chronic diseases, which can complicate care delivery ( 3 ). AI-enabled technologies such as predictive analytics, natural language processing, and decision-support systems offer opportunities to streamline clinical workflows, optimize resource allocation, and mitigate disparities in healthcare access ( 4 – 6 ). Studies have shown that AI-based clinical decision support tools can enhance diagnostic accuracy and reduce cognitive overload for clinicians, which is particularly beneficial in low-resource settings where providers often manage a wide range of complex cases with limited specialist support ( 7 ). Additionally, the use of AI-driven administrative automation, such as auto-populating clinical notes, processing billing, and managing electronic health records, can significantly reduce the time providers spend on non-clinical tasks ( 8 ). This is especially valuable in safety-net settings, where limited staffing and high caseloads often lead to documentation burdens that contribute to provider stress and burnout. A study incorporating specific AI models for Hispanic and Black women diagnosed with breast cancer shows how specific models can further improve diagnostic and treatment accuracy in underserved and underrepresented populations. These models outperformed general models by better predicting survival outcomes based on their unique health profiles. By tailoring AI models to the unique healthcare needs and disparities of specific demographic groups, these technologies can provide more accurate, individualized care, ultimately improving outcomes and promoting equity ( 9 ). For example, tests with Woebot and Crisis Textline chatbots showed a significant reduction in anxiety and symptom mitigation compared to historical values for human intervention ( 10 ). If these benefits can be confirmed, these tools could improve equal access, expand availability in terms of hours per day and user capacity, while offering anonymity and reduced stigma. Populations whose health suffers from the challenges of access and cost while simultaneously being underserved, uninsured, or have difficulty accessing care due to their location are safety net populations. These populations, including rural communities, have reduced access to quality health care and usually rely on publicly funded services because of economic hardship and systemic inequities while experiencing limited access to health care and social support. Although the specific factors may differ, structural barriers hinder their access to essential services and opportunities for improved health outcomes. Despite these advantages, health IT adoption in rural and underserved communities has historically lagged due to systemic barriers such as interoperability issues, digital infrastructure limitations, and financial constraints ( 2 , 11 ). Research consistently finds that one of the major barriers to advanced health-based technological solutions in rural areas is inadequate internet connectivity ( 12 – 16 ) And while lower-income households in urban areas have access to the internet primarily via mobile phones, they often encounter struggles to afford consistent access to broadband services ( 17 , 18 ). In both instances, the deployment of AI-based health solutions associated with remote patient monitoring or digital biomarkers face notable challenges. The implementation of AI in these settings must be approached with a focus on equity, feasibility, and long-term sustainability to avoid exacerbating existing healthcare disparities. As Brewer et al. (2020) highlight, AI models trained on predominantly urban and insured populations may fail to generalize to rural patient populations, leading to bias in diagnostic recommendations and clinical predictions. Addressing this issue requires deliberate efforts to ensure inclusive and representative training datasets that reflect the sociodemographic and epidemiological profiles of rural and underserved communities ( 19 ). A study in Tennessee incorporated social determinants of health (SDOH), such as access to healthcare, transportation, and socioeconomic status, to demonstrate the importance of tailoring AI models to local conditions. Machine learning models that integrated county-level SDOH data showed access to healthcare and vaccination rates were critical in predicting COVID-19 risk and could be applied similarly in rural settings to improve AI predictions and health outcomes. Incorporating community-specific data ensures that AI models are more accurate and reflective of the challenges faced by underserved populations, ultimately promoting equity in care delivery ( 20 ). Additionally, safety-net hospitals and rural clinics often lack the financial resources to invest in AI-driven infrastructure. Unlike for-profit institutions, these hospitals rely heavily on Medicaid and fixed tax revenues, which limits their ability to adopt new technologies without external funding or policy incentives ( 21 ). AI solutions for these settings must be cost-effective and easily integrated into existing electronic health records (EHRs) to maximize their impact. Browning et al. (2020) also emphasizes the need for vendor-neutral AI platforms that minimize interoperability challenges, as many safety-net hospitals operate with heterogeneous IT systems due to cost-driven procurement decisions. Furthermore, patient engagement and digital literacy remain critical factors influencing the success of AI-driven health interventions in rural areas. Studies indicate that racial and ethnic minority populations, who often comprise a significant portion of safety-net hospital patients, use digital health tools at lower rates due to barriers such as lack of trust in AI, poor access to broadband, and limited familiarity with health IT systems ( 7 ). A lack of digital literacy among patients and healthcare providers can limit the effective utilization of AI-driven tools ( 22 ). AI implementation strategies must therefore include culturally tailored patient education and community engagement efforts to ensure that underserved populations can meaningfully interact with these technologies. Methods Study Design Overview This study utilized a mixed methods approach to explore healthcare professionals’ perceptions of Health AI (HAI) technologies, particularly in the context of safety-net care. The research design included an extensive review of the literature, a health care provider survey, and a series of in-depth interviews with healthcare practitioners and researchers. These components allowed for the triangulation of insights across conceptual, experiential, and practice-based domains. The survey component focused on understanding current awareness and general attitudes toward Health AI, as well as reactions to four hypothetical AI tool deployment scenarios. Open-ended responses captured nuanced perspectives about AI’s utility, feasibility, and perceived risks across different clinical and patient contexts. The interviews were used to deepen understanding of themes identified in the literature and survey by eliciting rich, narrative accounts of participants’ real-world experiences and professional insights. All open-ended responses, including those tied to vignettes, were analyzed qualitatively using a shared coding framework. Study 1: Interviews To explore provider experiences and perspectives on health AI more deeply, the research team conducted 18 in-depth, semi-structured interviews. Fourteen interviews were conducted with healthcare practitioners, and four with research leaders. Participants were recruited through the research team’s professional network, including through referrals from health system partners. The interviews emphasized themes related to healthcare delivery in safety-net settings, with analytic focus placed on approximately six interviews where participants explicitly discussed rural, underinsured, or medically underserved populations. Interviews were conducted virtually via Zoom and lasted approximately 30 to 60 minutes. The interview guide included open-ended prompts that addressed a range of topics, such as: Levels of trust in AI and factors that enhance or erode that trust Perceived benefits and concerns surrounding AI integration in clinical workflows Ethical considerations and bias in AI models Infrastructure and interoperability challenges Patient-provider communication and engagement, especially in low-resource settings The role of AI in advancing or undermining health equity All interviews were transcribed verbatim and anonymized for analysis. We performed thematic analysis ( 14 ) with multiple coders. A qualitative coding framework consisting of 66 individual codes was developed to capture sentiments expressed across the interviews. These codes were grouped into broader thematic domains representing both the “promise” and “peril” of AI in healthcare. Selective coding was used to highlight particularly illustrative narratives related to safety-net implementation. Quotes from the interviews were later compared and aligned with themes that emerged from the survey data. Study 2: Surveys The second phase of data collection involved a statewide provider-facing survey designed to capture healthcare professionals’ perceptions of health AI, particularly in the context of rural and underserved populations. The survey was piloted in August 2024 within the research team’s healthcare provider network and subsequently fielded through Dynata, a healthcare panel provider, from September 5 to September 24, 2024. A total of 229 complete responses were collected from healthcare providers across Texas, representing a variety of roles and healthcare settings. While the majority of the survey consisted of quantitative items, several open-ended questions invited participants to elaborate on key topics. These questions were positioned directly after scaled questions and functioned as semi-structured qualitative data, offering context and rationale behind respondents’ ratings. The general open-ended questions included: “In a sentence or two, please describe what impacts your level of trust in AI technologies to support healthcare delivery.” (Linked to question which asked respondents to rate their current level of trust in HAI) “In a sentence or two, please explain why you feel this way about your patients’ response to AI to support better health outcomes.” (Linked to question which asked how responsive patients would be to HAI) “Please list two ways you think AI could specifically benefit rural and other safety-net populations in healthcare.” We performed thematic analysis ( 14 ) with multiple coders and have focused our analysis of emergent themes as appropriate by level of trust and the balance of appeal vs concern. In addition to these questions, the survey also included four health AI vignettes, using an approach similar to ( 23 ). Each vignette described a hypothetical AI tool for clinical use among safety-net populations. The technologies chosen for the vignettes were designed based on needs and interests expressed in the literature. For each vignette, respondents were asked to rate the tool’s appeal and potential concern on a Likert scale, followed by an open-ended prompt: “In a sentence or two, tell us a little more about the appeal you perceive and any concerns you have.” Summary descriptions for the four vignettes are: DocuScribeAI : A clinical documentation tool that transcribes and organizes provider-patient interactions in real time using natural language processing and integrates with electronic health records (EHRs). TherapiaAI : A tool for treatment personalization that analyzes data from patient histories, genetics, and current conditions to generate customized therapeutic plans within EHR systems. HealthRiskAI : A predictive analytics platform that incorporates social determinants of health (SDoH) to assess real-time patient risk and optimize care allocation for high-need individuals. CommCare : An AI tool that analyzes patient communication patterns to improve engagement and reduce no-shows in high-volume mental health clinics. Across the eight open-ended questions, including the three general prompts and the four vignette-based responses, nearly 200 qualitative responses were collected for each item. Responses were thematically coded using the same coding framework developed for the interview transcripts. Approximately 30 final codes were applied across the dataset, capturing provider perspectives on trust, ethical concerns, bias, implementation feasibility, cost, health literacy, provider benefit, and patient engagement. These codes were subsequently grouped into thematic domains to support interpretation of patterns across the entire dataset. Representative quotes from the survey are presented in the results section to illustrate key findings. Results Study 1: Narrative Insights from Thought Leader Interviews Our analysis of the exploratory interviews in Study 1 produced 5 major themes which we discuss next: Trust in AI (Community & Provider Perspective); Ethical Concerns, Bias, and Transparency; Financial and Integration Constraints in Safety-Net Contexts; Digital Literacy & Patient Engagement; and AI Integration Best Practices. Trust in AI (Community & Provider Perspective) Trust in AI is a recurring theme in both the literature and provider narratives, especially in safety-net contexts where patient-provider relationships are often shaped by systemic inequities, historical mistrust, and resource scarcity. For healthcare providers to feel confident in AI tools, they must believe in the reliability, clinical relevance, and transparency of algorithmic decision-making. But building trust goes beyond technical performance; it must also encompass cultural, emotional, and social dimensions, particularly for communities that have been underserved or harmed by healthcare systems in the past. As a chief operating officer of digital care platform explains, “Folks who are traditionally disadvantaged don't trust a lot of structure because they've felt abused, or they haven't felt that access... they have been turned down more than they have been accepted.” In her view, earning trust with these populations requires not just offering solutions, but forging a connection grounded in familiarity and empathy: You’ve got to align trust and comfort level with anybody who is using whatever we’re suggesting... not just point at people and say, ‘I’ve got this done for you.’ What is that human connection? What is that community and societal connection? This highlights the importance of culturally informed outreach and community-based engagement in building trust, not only in AI tools themselves, but in the systems deploying them. An executive leader in community health adds that trust is also shaped by demographic and contextual factors, noting that perceptions of AI may vary based on generational, cultural, and service-line considerations. She observes that “place of origin, culture, perspective are factors that are contributing to general trust in the system,” especially when working with underserved populations. For example, parents may be more skeptical of AI-supported pediatric care than for themselves, signaling how trust can be context-specific: Adults are like, ‘Sure, I’ll get on a telemedicine visit’... but ‘I want my kid to see the pediatrician, because I think the pediatrician needs to touch my kid. So, I can see that manifesting itself in AI as well. Finally, one faith-based mental health leader’s insight raises a critical concern about the validation and governance of AI tools, particularly in safety-net environments. For AI to be trusted, users must feel confident that there is oversight and accountability built into the system: Who’s doing the background check on AI?... It’s not a compensation for the lack of providers. We’re just trying to find a space where AI can support people in rural communities, because we can’t get to them. Together, these perspectives suggest that trust in AI must be earned on multiple fronts: technical reliability, cultural responsiveness, generational comfort, and institutional transparency. In safety-net settings, where providers are often the bridge between systems and communities, any AI implementation must be accompanied by careful attention to the social infrastructure that supports trust. Ethical Concerns, Bias, and Transparency Ethical concerns surrounding AI in healthcare are deeply rooted in questions of fairness, access, and unintended consequences. In safety-net settings, where the stakes are high and resources are limited, these concerns are magnified. Without deliberate efforts to address them, AI risks reinforcing longstanding inequities in health outcomes. An executive in rural healthcare leadership expresses a concern shared by many: that AI could exacerbate existing disparities by favoring large, well-funded institutions over smaller, rural providers. My biggest concern is that you're going to have two tiers of systems… Epic is built for large hospitals, and our smaller rural hospitals don't have access to it… If these tools… enhance the efficiency of care and quality of care… you're going to further widen the quality and care gap that exists today. His remarks highlight the ethical challenge of ensuring equitable access to AI tools, especially when those tools are designed with high-resource settings in mind. This observation illuminates a recurring concern about innovations in healthcare: that they tend to be built for more resource-rich healthcare providers, thus often overlooking the unique needs, challenges, and expertise of resource-poor healthcare providers. An expert in health care policy echoes this concern, drawing a parallel to the early days of telehealth. She points out that while large academic centers have the infrastructure to implement new platforms, rural settings often lack both governance and trust. “In rural settings… they don't have the governance structure… That’s a barrier because they don’t have the resources… and also trust… AI reminds me of how telehealth was 15–20 years ago… And I think that's where AI is going to be in a few years.” Her reflection signals the risk of repeating past mistakes in rolling out health technologies without equity-focused planning. A health data innovation leader emphasizes the dual potential of AI, to either narrow or widen disparities, depending on how it is implemented. “AI has the… opportunity to close the gap, but if we aren't trying to close the gaps, it will create a bigger gap… I live in a rural area, and it’s about… real, true access to health care.” Her comment reinforces the idea that ethical implementation is not automatic; it requires intentional strategies that prioritize greater attention to the context-specific needs of safety-net populations. Finally, an academic innovation researcher calls attention to the often-overlooked ways bias can creep into administrative systems. “AI in health system administration is probably under-observed… Hospitals may be unknowingly building biases into their administrative workflows.” This observation broadens the conversation about bias beyond clinical algorithms to the operational processes that structure how care is delivered. Together, these perspectives underscore that ethical AI implementation requires more than good intentions. It demands active attention to the ways that bias, access, and transparency shape outcomes and a commitment to designing systems that work for the communities most at risk of being left behind. Financial and Integration Constraints in Safety-Net Contexts The promise of AI in healthcare often clashes with the practical realities faced by safety-net institutions, where limited budgets, infrastructure gaps, and governance challenges present major barriers to adoption. In these contexts, financial constraints and technological fragmentation make implementation especially difficult. An executive in rural healthcare leadership highlights a core financial barrier: without immediate, demonstrable return on investment (ROI), rural hospitals are unlikely to take on the risk of adopting AI tools. Rural hospitals... would have to have an immediate ROI. A lot of them are cash strapped, as it is... you have to be able to show the savings to get them to buy into it, and then you still have the other barriers, like distrust and misunderstanding and all those other things. This participant’s comments reflect the pressure rural facilities face to justify every dollar spent, especially to decision-makers on local boards who may lack technical expertise or familiarity with digital health investments. A faith based mental health leader adds that technological infrastructure remains a major obstacle, particularly in geographically isolated areas with unreliable internet access. Those people who are in rural communities... they are also in heavily wooded, non-internet-savvy communities... We provide services in a rural school district... but we still don't have the technology capabilities for them to video into a counselor because the internet is so spotty. This leader’s example illustrates how integration challenges are not just about software compatibility, but also about physical connectivity and basic digital access. Together, these perspectives underscore that financial feasibility and technological readiness must be foundational considerations when designing and deploying AI in safety-net settings. Without affordable solutions and the infrastructure to support them, even the most promising tools will remain out of reach for the communities that could benefit most. Digital Literacy & Patient Engagement In rural and underserved communities, digital literacy and patient engagement are central to the successful adoption of AI tools. While the proliferation of smartphones and internet-enabled devices offers new opportunities to extend care, these tools must align with how people actually use technology, and more importantly, how they feel about using it. A chief operating officer of digital care platform emphasizes the need to build on existing comfort levels with everyday technologies like mobile phones. “No matter what else is going on in the world, they have a phone. So how do we seize that comfort level with AI? So how do we adapt to how people use technology?... it's how are we assuring that what we're developing truly will enhance who they are... and enable them to have better health outcomes... not only to care, but to food and to education and to resources.” Her framing suggests that AI implementation must start with a deep understanding of how people already engage with technology and how it can be adapted to support broader goals of well-being. A health equity executive (D.P) adds that digital literacy is not just about access but also trust and confidence in the systems being offered. “Our community health workers have to do a lot of one-on-one problem solving to help patients be comfortable with the technology... there's concerns about information security and quality of care. So, I think that's something that we need to work on with communities… is this a quality and secure alternative for them?” She highlights the labor-intensive effort often required to onboard patients into new systems, especially in communities where healthcare technology is still viewed with caution or skepticism. Together, these insights reinforce that engagement with AI in safety-net settings depends not only on access, but on familiarity, trust, and the perceived value of the technology. Tailoring solutions to community habits and building digital confidence are essential steps toward equitable and sustainable AI adoption. AI Integration Best Practices Successfully implementing AI in safety-net healthcare settings requires more than adopting the latest tools. It demands intentional design, context-aware planning, and structured collaboration between healthcare organizations and technology vendors. Best practices must be rooted in real-world challenges and informed by the operational complexities that providers face every day. A community health COO emphasizes the regulatory and practical realities of providing care in safety-net systems. “In a highly regulated industry like healthcare... the challenge is balancing face time with patients while handling documentation and regulatory requirements... those things that are in the world outside of the four walls of our clinic that impacts a patient’s ability to engage in their care... trying to mitigate those challenges on a day to day to day basis.” Her insight reflects how integration must consider not only the technology itself but also the broader workflow and compliance burdens that shape provider decision-making. A chief operating officer of digital care platform focuses on how AI can be used to better connect underserved patients with local services. She sees opportunity in AI’s ability to simplify access and personalize support: “If I’m an underserved person... how do we help that person in a respectful way, prioritize those community resources?... we can prove through artificial intelligence that we are actually having a positive social impact on community outreach and community services and healthcare.” Her emphasis on respectful guidance and local relevance suggests that integration efforts must be attuned to the everyday realities and needs of the communities they serve. - Worker in Healthcare. It is perspectives like these that are informed by engagement with safety-net communities and familiarity with their needs that make this specific group of practitioners crucial to the design of relevant AI solutions in contexts like these. An executive in rural healthcare leadership adds that governance and vendor collaboration are critical, particularly in under-resourced systems that may lack internal capacity for AI oversight. “Most of the rural hospitals are ignorant of the capability of the tools... I do think they need some help... maybe having the vendors create a template and having the internal team discuss it... what is missing from this? What other considerations do we need to think about?” His comments point to the importance of co-design and shared accountability between technology providers and health systems. Together, these perspectives suggest that successful AI integration must involve thoughtful workflow alignment, community-centered design, and collaborative governance structures. Integration isn’t just about embedding AI into clinical systems; it is also about building the trust, infrastructure, and organizational processes that make it work in practice. Study 2: Thematic Analysis of Provider Perceptions about AI in Healthcare As described under Methodology above, Study 2 included semi-structured lines of questioning about how participants’ thought about: Their current level of trust in AI technologies to support healthcare delivery Their perceptions about their patients’ response to AI to support better health outcomes The ways they thought AI could specifically benefit rural and other safety-net populations in healthcare Our analysis of the first 3 topics includes an overall view of the sample plus a stratification by current level of trust (low vs high). Current Level of Trust Table 1 presents the themes that emerged from a thematic analysis of open-ended survey data ( 24 , 25 ). It shows the frequency of the various themes influencing participants’ degree of trust in AI technologies for health care delivery among low trust and high trust groups. Table 1 Issues Impacting Trust in AI in Health Care Total Sample Low trust High trust Sig Building trust 39.8% 50.5% 28.4% 0.001 Data bias and accuracy 25.6% 33.0% 17.6% 0.011 Maintaining data security, privacy, & anonymity 13.7% 12.8% 14.7% Efficient and effective care 13.3% 3.7% 23.5% 0.000 Adequate policy, governance, & oversight 9.0% 8.3% 9.8% Administrative tasks 8.1% 2.8% 13.7% 0.003 Diagnostic capabilities 7.1% 2.8% 11.8% 0.011 Need for human in the loop 7.1% 6.4% 7.8% Accuracy of data 6.6% 1.8% 11.8% 0.004 Automate documentation 6.6% 2.8% 10.8% 0.019 Personalized care 5.2% 7.3% 2.9% Desire for human touch 4.7% 6.4% 2.9% Decision making 4.3% 4.6% 3.9% Questioning if AI is needed 4.3% 7.3% 1.0% 0.022 Interplay between AI and user 3.8% 1.8% 5.9% Enhanced health information / education access 2.8% 2.8% 2.9% Changing provider roles 2.4% 3.7% 1.0% Cost and hesitancy to change 1.4% 1.8% 1.0% Governance structure 1.4% 1.8% 1.0% Time savings 1.4% 0.9% 2.0% Care for the marginalized 0.5% 0.0% 1.0% Constant need to adapt to trends 0.5% 0.0% 1.0% Ethical AI 0.5% 0.0% 1.0% Health literacy and trust 0.5% 0.0% 1.0% Interplay between language, culture, & physicality 0.5% 0.0% 1.0% Social determinants of health 0.5% 0.0% 1.0% Workforce availability, training, & turnover 0.5% 0.0% 1.0% Other 23.7% 16.5% 31.4% 0.011 Total 211 109 102 Caption : Table 1 presents the themes that emerged from a thematic analysis of open-ended survey data. The frequency of the various themes influencing participants’ degree of trust in AI technologies for health care delivery are shown for the total and among low trust and high trust groups. Trust differences that were significant at p < = 0.05 are noted. As shown in Table 1 , the degree of trust in AI among low-trust respondents and high-trust respondents are influenced by different factors. Among low-trust respondents, over half (50.5%) cited building trust as a core issue, followed closely by data bias and accuracy (33.0%) and maintaining data security, privacy, and anonymity (12.8%). These participants expressed deep concern over the fairness, reliability, and safety of AI systems. In contrast, high-trust respondents were more likely to emphasize the practical benefits of AI: 28.4% mentioned building trust, but significantly more cited efficient and effective care (23.5%), diagnostic capabilities (11.8%), and automating documentation (10.8%) as reasons for their confidence in AI. Additionally, high-trust respondents were far more likely to mention the value of adequate governance and accuracy of data compared to their low-trust counterparts. Both groups identified building trust as the most frequently mentioned factor behind their current level of trust. (Note the high trust group includes a majority of respondents stating they have a trust level of “somewhat”.) High trust respondents were more likely to emphasize the importance of ( 1 ) data bias and accuracy; ( 2 ) adequate policy, governance, and oversight; and ( 3 ) administrative tasks. They also value AI's ability to automate documentation, support diagnostic capabilities, and provide efficient and effective care. Maintaining data security, privacy, and anonymity also significantly influences trust among this group. These findings suggest that individuals with high trust focus on AI's ability to streamline processes; enhance governance; and deliver secure, efficient, and effective outcomes. For respondents with lower trust levels, the concerns are more targeted, with two key factors emerging as significant - data bias and accuracy are central concerns as well as maintaining data security, privacy, and anonymity - indicating that low trust respondents are particularly sensitive to the fairness and privacy of AI systems. Additionally, ensuring adequate policy, governance, and oversight significantly impacts trust for this group, reflecting fears around the potential misuse or breaches of sensitive information. In the course of our analysis of the qualitative data about trust, four themes emerged, which we identify and discuss below. Building Trust A central theme that emerged across responses was the foundational role of trust when integrating AI into healthcare. Providers consistently expressed skepticism, largely due to the relative novelty of the technology, the absence of long-term outcome data, and the ongoing need for human oversight. A nurse in the high-trust group described AI as “an emerging technology, one that still needs to get the bugs worked out,” reflecting early-stage hesitation. Similarly, a graduate student healthcare practitioner remarked, “There are not enough studies with long-term outcomes to prove AI is safe and effective,” pointing to the limited empirical validation of AI tools in clinical practice. Even among those open to its potential, there was strong consensus on maintaining human authority in medical decision-making. One nurse firmly stated, “AI in no way should diagnose or recommend any medical guidance,” emphasizing that clinical decisions should remain with licensed professionals. A low-trust respondent echoed this need for caution, stating, “It is very new and more studies regarding this need to be done,” underscoring how trust remains contingent on evidence and transparency, particularly for those less familiar or confident in the technology. In some clinical situations like radiology, for example, AI's ability for diagnostics is the equivalent to or even better than human professionals. Still, human experts are required to confirm diagnosis, communicate the diagnosis to colleagues, and deploy emotional intelligence and context awareness in working with patients to develop a viable treatment plan ( 26 ). Data Bias and Accuracy Concerns about data bias and accuracy were also prominent, particularly among those who expressed conditional trust in AI. A physician explained, “I trust that AI technology can sift through data... but I do not trust the results 100% since there are many examples of either confabulation or incorrect clinical identification,” highlighting the tension between data processing capabilities and clinical reliability. These concerns were even more pronounced among low-trust respondents. One physician specifically pointed out, “The data AI technology is using is not diverse and inclusive,” underscoring the risks of algorithmic bias and the need for datasets that reflect diverse patient populations to ensure equitable care. As we develop deeper knowledge of health AI systems and how to deploy them effectively, the definition of diverse patient populations will require additional thought. For example, the deployment of AI in a rural setting will need to be highly sensitive to the context specific features and lived experiences that define life in a particular rural setting. This means that developing algorithmic solutions that are context aware and culturally sensitive will require approaches to governance that are not fully established. Maintaining Data Security, Privacy, and Anonymity Maintaining data security, privacy, and anonymity emerged as another significant determinant of trust, with concerns spanning both high- and low-trust groups. A nurse noted, “The recent media about AI compromising people’s privacy online makes me a bit wary,” capturing a broader apprehension around digital vulnerability. A physician added, “My biggest concern with AI would be the risk of getting hacked,” pointing to fears around system breaches and data misuse. Even among high-trust participants, the need for formal safeguards remained essential. As one physician emphasized, “I need to know that the AI is going to be HIPAA compliant. The platform it is on also impacts whether I feel safe with it.” Together, these perspectives signal that trust is inseparable from strong data governance and transparency around privacy protections. Efficient and Effective Care High-trust respondents more often emphasized the efficiency and effectiveness of AI in improving healthcare delivery. A nurse described AI as “very effective and sufficient in aiding part of our medical system and staff,” illustrating its perceived ability to streamline operations. A physician supported this view, citing emerging research that suggests AI “may help throughput,” especially in areas like diagnostic imaging. Another physician explained how AI could “integrate records and abstract (extract?) important information,” noting its utility in managing documentation, generating differential diagnoses, and supporting treatment planning. However, low-trust perspectives reveal concern that the pursuit of efficiency may come at a cost to patient care. As one physician cautioned, “Feel like there is going to be a push from higher ups (whether administration or Medicare/Medicaid) to use AI to cut costs and it will create problems in medicine,” highlighting that perceived administrative motives may undermine trust in AI’s implementation, particularly when efficiency is prioritized over clinical judgment. Perceived Benefits of Health AI for Safety Net and Rural Populations Participants identified a variety of ways in which AI could positively impact care delivery for underserved and rural communities. Table 2 summarizes the most frequently mentioned themes across trust groups, offering a comparative view of how low- and high-trust respondents envision AI contributing to safety-net care. Table 2 Potential Perceived Benefits of Health AI for Safety Net and Rural Populations Total Low trust High trust Sig. Efficient and effective care 23.5% 22.4% 24.5% Administrative tasks 19.4% 17.3% 21.4% Decision making 16.8% 19.4% 14.3% Care for the marginalized 11.7% 12.2% 11.2% Diagnostic capabilities 11.7% 11.2% 12.2% Personalized care 11.2% 8.2% 14.3% Enhanced health information / education access 9.2% 11.2% 7.1% Lack of time, funds, or resources 6.6% 7.1% 6.1% Telehealth/telemedicine 6.1% 1.0% 11.2% 0.003 Social determinants of health 4.1% 4.1% 4.1% Automate documentation 3.6% 2.0% 5.1% Interplay between AI and user 3.6% 5.1% 2.0% Constant need to adapt to trends 2.0% 3.1% 1.0% Data bias and accuracy 1.0% 1.0% 1.0% Health literacy and trust 1.0% 1.0% 1.0% Intense and regulated industry 1.0% 1.0% 1.0% Changing provider roles 0.5% 1.0% 0.0% Cost and hesitancy to change 0.5% 1.0% 0.0% Maintaining data security, privacy, & anonymity 0.5% 1.0% 0.0% Other 68.4% 67.3% 69.4% Total 196 98 98 Caption : Table 2 presents the themes that emerged from a thematic analysis of open-ended survey data. The frequency of the various themes influencing participants’ perceptions about patients’ response to AI to support better health outcomes shown for the total and among low trust and high trust groups. Trust differences that were significant at p < = 0.05 are noted. Table 2 depicts how low trust and high trust respondents perceive the potential benefits of AI for rural and safety net services populations. Among all respondents, “efficient and effective health care” emerged as the leading Health AI benefit. For low trust respondents, “decision making,” 38%, was the second leading benefit, followed by “administrative tasks.” For high trust respondents, “administrative tasks,” 32%, was the second leading benefit, and “personalized care” was third. Telehealth/telemedicine was significantly more emphasized by high trust respondents (11%) compared to low trust respondents (1%), highlighting that individuals who trust AI technologies may recognize the potential of AI-powered telehealth solutions to bridge gaps in service delivery and enhance care availability for remote or resource-limited populations. Efficient and Effective Care (Top code for both trust groups) Across both high and low trust groups, the most frequently cited benefit of health AI was its potential to deliver more efficient and effective care. A physician in the high trust group noted that AI “has the potential to lead to more comprehensive patient care and reduce wait times,” reflecting confidence in its ability to improve throughput and patient satisfaction. Similarly, an administrative staff member observed, “Because I think that when they see how good it is, they will become a fan,” suggesting that positive experiences with AI could drive wider patient acceptance. A nurse echoed this sentiment, explaining, “Patients want care and they want it fast. They want answers and to feel better. AI would assist in all of these things,” highlighting how timeliness and responsiveness are central to patient needs and may be supported by AI-driven tools. By contrast, a low-trust physician noted, “patients would perceive it as a benefit only, if a noticeable care change affects them.,” underscoring the conditional nature of acceptance in settings where skepticism remains high. Administrative Tasks (Emphasized more by high trust group) Administrative burden reduction was another key theme, particularly among high trust respondents. The use of AI-based automative systems to support administrative tasks has been one of the earliest and most frequent applications in healthcare ( 8 , 27 , 28 ). Tasks that are routine in nature have been most amenable to automation. Moreover, the significant time expenditure on administrative tasks like clinical notes, billing, and management of electronic health records contributes to high levels of stress and burnout among healthcare practitioners. A nurse practitioner stated that “AI is likely an increasing tool in many services to include healthcare. It may help to streamline care and reduce administrative tasks,” suggesting that automation could help free up provider time for direct patient care. An administrative staff member similarly emphasized the benefits of convenience and access, explaining, “I’m sure they will feel better because they will not have to be struggling and they can get the records very very easy through the technology.” An allied health professional added that AI “can create better connections and source for administrative staff,” reinforcing the idea that intelligent systems can enhance internal operations and reduce inefficiencies. One school-based nurse in the low trust group expressed caution, stating, “I'm not exactly sure how AI fits into nursing care in a public-school setting. It's a busy place and I can see the appeal of AI to help with administrative tasks, but privacy concerns outweigh the potential benefit, in my opinion.” This view illustrates that for some low-trust respondents, the perceived risks, particularly around privacy, may still overshadow the administrative advantages of AI. Decision Making (Emphasized more by low trust group) Decision support was cited more frequently by low trust respondents, indicating that while they recognized AI’s potential to assist with clinical judgments, they remained cautious. A member of clinical staff in the low trust group shared that AI could be useful for “understanding and summarizing the clinical issues with the client,” pointing to its capacity for synthesizing complex data. However, others expressed reservations: a nurse practitioner emphasized the need for more evidence and security, stating, “More information [is] needed to improve confidence in AI. Plus, there remains a lack in tight electronic information security.” A physician echoed these concerns, warning that “they would need to be informed that using AI comes with risks.” These responses reveal that while decision-making support is valued, it must be accompanied by safeguards and transparency to be embraced in lower trust environments. By contrast, high trust respondents conveyed greater confidence in AI’s role in improving care through better decision-making. One provider stated, “AI has the potential to significantly improve the precision and effectiveness of healthcare. By analyzing vast amounts of data, AI can help in early detection of diseases, personalize treatment plans, and provide continuous monitoring and support.” Personalized Care (Emphasized more by high trust group) Personalized care was more commonly emphasized by high trust respondents, who viewed AI as a way to enhance, not replace, individualized medicine. A nurse practitioner explained, “They would appreciate quicker care but still want personalized care,” recognizing the dual importance of speed and customization. Another nurse practitioner added that “some patients, especially under 40, may better understand and appreciate the personalization,” pointing to generational differences in comfort with technology. Even among low trust respondents, there was some openness to this potential, with an administrative staff member noting, “It will create an easier process if things are artificially modified to what actually fits every patient’s needs.” These responses reflect cautious optimism that AI, when properly tailored, can support more patient-centered approaches. Telehealth/Telemedicine (Significantly more emphasized by high trust group) Telehealth and telemedicine were significantly more emphasized by high trust respondents, reflecting greater enthusiasm for technology-enabled remote care. A clinical staff member noted, “We use AI in our everyday use now, and in most cases have had a better experience. So why would we not want that for our healthcare?” Their statement suggests normalization of AI in other aspects of life has helped lay the groundwork for acceptance in medical contexts. A physician similarly observed, “With technology today it’s more acceptable for patients to try new [tools],”- (R_3Ey20tPbO4KkwgZ) underscoring how digital familiarity can support the adoption of AI-powered telehealth solutions. However, caution remained among some low trust providers; one physician commented, “Patients are open to being seen sooner and getting the services they need in a timely manner. However, they need to be informed that using AI comes with risks.”- (R_5vZXZ0W1BN8eM4y) This remark highlights a recurring theme throughout the study: trust must be earned through transparency, validation, and meaningful patient education. Health AI Benefits for Rural and other Safety-Net Populations To better understand how healthcare professionals perceive the impact of AI on rural and other safety-net populations, respondents were asked to identify specific ways AI could improve care delivery in these contexts. These insights provide a more grounded view of how AI might be integrated into settings marked by resource constraints and patient vulnerability. Table 3 summarizes the organizational factors that providers believe will shape the feasibility and sustainability of Health AI implementation in these environments. Table 3 Considerations for Organization Adoption of Health AI Total Low trust High trust Sig. Cost and hesitancy to change 23.9% 38.5% 13.2% ) Interplay between AI and user 16.3% 20.5% 13.2% Maintaining data security, privacy, & anonymity 13.0% 12.8% 13.2% Data bias and accuracy 9.8% 7.7% 11.3% Workforce availability, training, & turnover 9.8% 5.1% 13.2% Adequate policy, governance, & oversight 7.6% 2.6% 11.3% Lack of time, funds, or resources 7.6% 10.3% 5.7% Unequal access to technology/ infrastructure 6.5% 12.8% 1.9% 0.036 Changing provider roles 5.4% 5.1% 5.7% Building trust 4.3% 5.1% 3.8% Desire for human touch 4.3% 7.7% 1.9% Administrative tasks 2.2% 2.6% 1.9% Automate documentation 2.2% 0.0% 3.8% Constant need to adapt to trends 2.2% 2.6% 1.9% 0.005 Efficient and effective care 2.2% 0.0% 3.8% Enhanced health information / education access 2.2% 2.6% 1.9% Intense and regulated industry 2.2% 2.6% 1.9% Questioning if AI is needed 2.2% 2.6% 1.9% Chronic health condition prevention and treatment 1.1% 0.0% 1.9% Decision making 1.1% 2.6% 0.0% Diagnostic capabilities 1.1% 0.0% 1.9% Other 35.9% 35.9% 35.8% Total 92 39 53 Caption : Table 3 presents the themes that emerged from a thematic analysis of open-ended survey data. The frequency of the various themes influencing participants’ adoption and deployment of health artificial intelligence for health care delivery are shown for the total and among low trust and high trust groups. Trust differences that were significant at p < = 0.05 are noted. Table 3 highlights additional factors that respondents believe are relevant to their organization's adoption and deployment of Health AI. Cost and hesitancy to change was significantly more emphasized by low trust respondents, reflecting their concerns about financial barriers and organizational resistance to AI integration. On the other hand, unequal access to technology/infrastructure was significantly more emphasized by high trust respondents, indicating that those with greater confidence in AI are more focused on addressing systemic disparities in resource availability to ensure successful implementation. These findings illustrate that addressing financial and cultural resistance is key for low trust groups, while high trust groups prioritize infrastructure equity as a critical enabler of AI adoption. Cost and Hesitancy to change Cost and hesitancy to change emerged as key concerns, particularly among low-trust respondents. Providers in this group expressed doubts about the feasibility of implementing AI given financial constraints and the potential for disruption to patient-provider dynamics. One nurse emphasized that adoption would be more acceptable “if it could be implemented with limited intrusion into the provider and patient relationship and did not require a robust electronic record infrastructure or expensive funding to maintain.” A physician echoed this sentiment, acknowledging AI’s potential benefits but noting “significant front facing costs,” which may deter resource-constrained organizations from exploring these tools further. Even among high-trust respondents, financial burden remained a cautious undercurrent; one physician remarked, “I think it can decrease resource needs (with significant front facing costs),” highlighting that while optimism exists, concerns about initial investment persist across trust levels. Interplay between AI and user Another important theme was the interplay between AI and the end user, including both patients and providers. For low-trust respondents, ease of use was critical. One nurse suggested, “Make it very basic so patients will feel comfortable with AI,” underscoring concerns about accessibility and digital literacy. In contrast, high-trust respondents viewed user familiarity as a strength. An allied health professional observed that many patients already “know about AI and support it,” suggesting that acceptance may be growing in some communities. A healthcare IT professional added that AI integration may help with workforce recruitment, stating that “AI can improve the provider experience with EMR and help us recruit the newer crop of providers that are learning about AI in school.” These perspectives indicate that the user experience, both in terms of design simplicity and professional training, will significantly shape adoption trajectories. Data security, privacy, and anonymity Data security, privacy, and anonymity were persistent concerns across both trust groups. A member of administrative staff in the low-trust group noted that while AI might be helpful for tasks like translation, “privacy-wise it’s a bit scary.” A clinical staff member, also from a low-trust group, reinforced the importance of this issue: “Maintaining data security and privacy will be extremely important for patient trust.” At the same time, a high-trust clinical provider pointed to AI’s potential to advance equity, stating, “Possibly the first and/or most neutral approach to healthcare, removing prejudice, racism, sexism, etc... from the equation.” These comments highlight the tension between AI’s transformative potential and the ethical guardrails required to protect patient data and reinforce public trust, particularly in vulnerable communities. Study 2: Thematic Analysis of Provider Reactions to Specific Health AI Tools The survey included four vignettes about four hypothetical AI tools, designed to assess the appeal and concern for specific Health AI concepts and technologies. The vignettes presented the opportunity to move from abstract to more specific applications of AI in healthcare. See Appendix for full vignettes, which can be summarized as follows: DocuScribeAI is an AI tool designed to assist in transcribing and organizing clinical notes during patient interactions. TherapiaAI analyzes real-time data from patient histories, genetic information, and current health conditions. It suggests customized therapeutic plans based on clinical guidelines and patient-specific factors. HealthRiskAI integrates patient-specific social determinants of health and provides real-time risk assessments. CommCare analyzes communication preferences and engagement patterns, offering insights to improve patient-provider interactions. DocuScribe draws from AI-powered transcription technologies, particularly large language models (LLMs), to automate clinical documentation and ease administrative burden. These tools generate structured notes from physician–patient interactions, improving workflow efficiency and reducing burnout, while enabling more patient-centered care. However, risks such as omissions, fabrications, misattributions, and limited contextual understanding, especially around nonverbal cues and speaker differentiation, can compromise accuracy ( 29 ). TherapiaAI is inspired by AI tools in personalized medicine that use genomic, clinical, and lifestyle data to generate individualized treatment plans. These applications enhance diagnostic precision and enable earlier interventions, especially in complex or ambiguous cases, by synthesizing diverse health inputs. Yet, they also raise concerns around biased data, privacy of sensitive genomic information, and clinician challenges in interpreting complex AI outputs due to limited training and infrastructure ( 30 ). HealthRiskAI builds on AI systems that incorporate patient-specific social determinants of health (SDOH) to improve diagnosis and care navigation ( 20 ). These models dynamically integrate factors such as poverty, transportation access, and vaccination rates to produce real-time risk forecasts and guide equitable care coordination. For these systems to fulfill their promise, designers must ensure bias mitigation, interoperability of data, and transparency in how AI-generated risk insights are used in clinical workflows ( 31 ). CommCare was inspired by AI tools that analyze communication preferences and engagement behaviors to improve patient-provider interactions. Using data from EHRs, portals, and care platforms, these tools identify how patients prefer to receive information (e.g., SMS vs. calls) and can flag early signs of disengagement, prompting proactive outreach ( 32 ). By aligning communication strategies with patient-specific behaviors, these systems enhance adherence, satisfaction, and trust in care relationships ( 33 ). Appeal and concern were measured on 5-point scales, as shown next. Appeal 1) No Appeal 2) A Little Appeal 3) Some Appeal 4) Highly Appealing 5) Extremely Appealing Concern 1) No Concern 2) A Little Concerning 3) Some Concern 4) Highly Concerning 5) Extremely Concerning Figure 1 shows the mean appeal and concern rating for each vignette, including the 95% confidence interval (CI). Across all AI concepts/tools, the average appeal is greater than the average level of concern. The appeal of CommCare is the lowest and is significantly lower than the most appealing concept, DocuScribeAI. Caption: Fig. 1 displays the mean scores (plus 95% confidence intervals) for the appeal and concern ratings for the four concepts examined in our survey. Participants were also asked: “In a sentence or two, tell us a little more about the appeal you perceive [for each fictional AI tool] and any concerns you have.” Responses were analyzed thematically and coded. Tables 4 through 7 show these responses for each vignette, with the results stratified by participants as follows: Those labeled Promise in the tables and our discussion of our results gave an appeal rating greater than or equal to their concern rating. Those labeled Peril gave a concern rating greater than their appeal rating. Our analysis of the perceptions about the 4 specific, hypothetical Health AI technologies includes an overall view of the sample plus a stratification based on whether perceptions of appeal were greater than perceptions of concern (promise vs. peril). DocuScribeAI Data from Vignette 1 highlights that, while DocuScribeAI is widely recognized for its ability to automate documentation and save time, skepticism remains about its accuracy and security, and the need to maintain human oversight. The tool's potential to streamline administrative tasks and improve efficiency is promising yet concerns about data reliability and privacy must be addressed to build trust among skeptical stakeholders. Balancing automation with human involvement and ensuring a governance network is in place are critical to maximizing the perceived benefits while mitigating risks. There were no statistically significant differences in findings about this tool among high trust and low trust groups. See Table 4 . Data from Vignette 1 highlights that DocuScribeAI is widely perceived as a valuable tool for automating documentation (49.5% high trust, 38.1% low trust) and saving time (32.0% high trust, 32.4% low trust). Respondents also frequently noted the tool’s potential for improving accuracy of data (38.1% high trust, 32.4% low trust) and reducing administrative burden. However, some concerns remain, particularly among low trust respondents, regarding the need to maintain human oversight (26.8% high trust, 32.4% low trust), ensure data security and privacy (14.4% high trust, 11.4% low trust), and validate whether AI is appropriate or even needed in clinical workflows (7.2% high trust, 14.3% low trust). Table 4 DocuScribeAI: Reasons for Appeal and Concern Total Peril Promise Sig. Automate documentation 43.7% 21.4% 47.4% 0.010 Accuracy of data 35.2% 46.4% 33.3% Time savings 32.2% 14.3% 35.1% 0.029 Need for human in the loop 29.6% 46.4% 26.9% 0.036 Maintaining data security, privacy, & anonymity 13.1% 14.3% 12.9% Questioning if AI is needed 11.1% 21.4% 9.4% 0.059 Building trust 10.1% 10.7% 9.9% Efficient and effective care 10.1% 7.1% 10.5% Adequate policy, governance, & oversight 8.0% 10.7% 7.6% Interplay between AI and user 5.5% 10.7% 4.7% Interplay between language, culture, & physicality 5.5% 14.3% 4.1% 0.029 Administrative tasks 3.0% 7.1% 2.3% Changing provider roles 3.0% 3.6% 2.9% Workforce availability, training, & turnover 3.0% 0.0% 3.5% Decision making 2.5% 3.6% 2.3% Cost and hesitancy to change 2.0% 0.0% 2.3% Data bias and accuracy 2.0% 0.0% 2.3% Constant need to adapt to trends 1.0% 0.0% 1.2% Desire for human touch 1.0% 3.6% 0.6% Diagnostic capabilities 1.0% 0.0% 1.2% Personalized care 1.0% 3.6% 0.6% Lack of time, funds, or resources 0.5% 3.6% 0.0% Telehealth/telemedicine 0.5% 0.0% 0.6% Unequal access to technology/ infrastructure 0.5% 0.0% 0.6% Other 4.0% 0.0% 4.7% Total 199 28 171 Caption : Table 4 presents the themes that emerged from a thematic analysis of open-ended survey data. The frequency of the various themes influencing participants’ perceptions about promises and perils of DocuScribeAI are shown for the total and among promise and peril groups. Appeal and concern differences that were significant at p < = 0.05 are noted. Although both groups recognize the potential benefits of increased efficiency (12.4% high trust, 7.6% low trust), the presence of skepticism, especially related to governance and unintended consequences, highlights the importance of transparent oversight frameworks (6.2% high trust, 9.5% low trust). 1. Automate Documentation For respondents in the Promise group, DocuScribeAI was overwhelmingly viewed as a time-saving solution to the documentation burden that plagues many healthcare professionals. A physician praised its integration and efficiency, explaining that the tool “transcribes in real time and [is] integrated with EMR,” and noted it “should save time if it does not require extensive editing.” Echoing this, a nurse shared enthusiasm for the tool’s ability to remove a frustrating task from clinical workflows, stating, “That takes medical transcribing off the table and that’s wonderful [for] all hospital staff.” The reduction in burnout was also a recurring theme, with another physician observing, “Providers hate spending so much time on the documentation, that is a major cause of their day-to-day stress. Anything which can alleviate it will decrease burnout.” However, among the Peril group, enthusiasm was tempered by skepticism about accuracy and workflow disruption. One nurse questioned the tool’s necessity altogether, asking, “Charting isn't so overwhelming that AI is needed. How much time would I spend correcting what AI did??” This sentiment underscores the concern that rather than saving time, the technology could introduce new inefficiencies if not carefully implemented. 2. Time Savings Beyond automating documentation, promise-oriented respondents emphasized time savings as a key appeal. One physician framed the tool as a way to improve efficiency, remarking, “It could save massive amounts of time allowing [providers] to see patients more effectively,” though they cautioned that this benefit depends on the tool’s accuracy. A nurse succinctly noted, “Decreased paper time,” while a nurse practitioner elaborated that DocuScribeAI could “streamline documentation, allow more time to engage patients, allow providers to see more patients, complete documentation in clinic, [and have] no need to take work home.” These responses reflected an optimistic view that, if accurate, the tool could support higher-quality patient care while also improving provider work-life balance. However, not all respondents shared this confidence; one provider in the Peril group expressed skepticism, stating, “I'm concerned that it would take as much time to proofread and edit the AI results as it does to just chart myself,” highlighting fears that promised time savings might not materialize in practice. 3. Accuracy of Data Still, even among those in the Promise group, concerns about accuracy were not absent. A nurse practitioner admitted some skepticism, saying, “I’m concerned that it would take as much time to proofread and edit the AI results as it does to just chart myself.” Another nurse expressed concern over potential clinical risk, cautioning that “mistakes are made, words or medications [could be] heard incorrectly, which could cause major issues if not caught in a timely manner.” A physician from the Peril group echoed this hesitation, noting, “It sounds very appealing, but I would be worried about the accuracy of the transcription and documentation.” Despite one participant describing it as “nice to be able to go back and review those notes and know that they’re likely to be very accurate,” the Promise group emphasized that speed alone would not justify the use of the tool if accuracy could not be ensured. 4. Need for Human in the Loop The importance of maintaining a human in the loop was central across both Promise and Peril perspectives. For some, this was a manageable requirement. A nurse in the Promise group raised a critical point about accountability, asking, “What errors will there be anli8od who does the weight of the error fall on?” This concern becomes more pointed among those in the Peril group. A nurse stated, “I think it would have to be reviewed for accuracy before completely trusting something like this,” suggesting that confidence in AI-generated notes remains fragile without reliable safeguards. A researcher went further, pointing out that “you still need to approve the notes. May not even make sense,” illustrating a fear that the tool could introduce confusion or workflow delays if its outputs are not clinically usable without intervention. TherapiaAI Table 5 highlights perceptions of TherapiaAI, a tool designed to enhance treatment personalization. While both low trust and high trust groups emphasized the need for human oversight (19.5% low trust,17.4% high trust) and data bias/accuracy (18.4% low trust, 14.0% high trust), their underlying priorities diverged. Low trust respondents were significantly more focused on building trust (16.1%) and data security, privacy, & anonymity (9.2%), underscoring a need for transparency and protection to build confidence in the tool. In contrast, high trust respondents placed greater emphasis on personalized care (23.3%) and efficient and effective care (22.1%), suggesting optimism about AI’s role in improving patient outcomes. Both groups also cited enhanced health information and education access (13.8% low trust, 12.8% high trust) and time savings (6.9% low trust, 9.3% high trust) as potential benefits, but differences in emphasis reveal that trust level shapes how the benefits of AI are perceived. Table 5 TherapiaAI: Reasons for Appeal and Concern Total Peril Promise Sig. Need for human in the loop 18.5% 20.8% 18.1% Personalized care 18.5% 20.8% 18.1% Data bias and accuracy 16.2% 25.0% 14.8% Efficient and effective care 16.2% 4.2% 18.1% 0.085 Enhanced health information / education access 13.3% 8.3% 14.1% Building trust 10.4% 16.7% 9.4% Maintaining data security, privacy, & anonymity 8.7% 16.7% 7.4% Time savings 8.1% 4.2% 8.7% Desire for human touch 5.8% 16.7% 4.0% 0.014 Diagnostic capabilities 4.0% 8.3% 3.4% Changing provider roles 3.5% 8.3% 2.7% Decision making 3.5% 4.2% 3.4% Accuracy of data 2.9% 4.2% 2.7% Administrative tasks 2.9% 4.2% 2.7% Cost and hesitancy to change 2.3% 8.3% 1.3% 0.034 Automate documentation 1.7% 4.2% 1.3% Constant need to adapt to trends 1.7% 4.2% 1.3% Interplay between AI and user 1.7% 8.3% 0.7% 0.008 Questioning if AI is needed 1.7% 4.2% 1.3% Governance structure 1.2% 0.0% 1.3% Adequate policy, governance, & oversight 0.6% 0.0% 0.7% Care for the marginalized 0.6% 0.0% 0.7% Other 29.5% 16.7% 31.5% Total 173 24 149 Caption : Table 5 presents the themes that emerged from a thematic analysis of open-ended survey data. The frequency of the various themes influencing participants’ perceptions about promises and perils of TherapiaAI are shown for the total and among promise and peril groups. Appeal and concern differences that were significant at p < = 0.05 are noted. 1. Need for Human in the Loop Respondents uniformly expressed a strong commitment to maintaining a “human in the loop” when deploying TherapiaAI. A nurse from the promise group emphasized that “there should always be a human reviewing or monitoring AI outputs; machines can’t replace clinical judgment and compassion,” capturing the belief that no matter how advanced the technology becomes, the empathetic, nuanced insight of a clinician remains irreplaceable. This perspective was echoed by a physician from the promise group who stated, “I am concerned about AI being misaligned with my values and making decisions without provider involvement,” while another nurse reinforced, “Even with smart tools, oversight is essential. We can’t just rely on algorithms and hope for the best.” One administrative staff member from the peril offered a pointed caution: “There’s a potential risk that clinical staff might rely too heavily on this tool, potentially neglecting their own clinical and personal judgment. To prevent this, significant safeguards must be established to ensure that the AI complements rather than replaces the valuable insights and decisions made by healthcare professionals.” 2. Personalized Care Personalized care emerged as another critical advantage, especially from the promise perspective. A graduate student healthcare practitioner highlighted the transformative potential of the tool by noting, “AI can tailor care to each individual’s needs, which is especially important for marginalized populations who are often overlooked.” This sentiment was shared by a nurse who observed that “with AI, patients might finally receive recommendations that truly reflect their history, preferences, and goals.” Further reinforcing this promise, another nurse commented that “personalized care is essential to building trust and better outcomes, and AI can help us get there if used carefully.” Still, not all respondents were fully convinced - one nurse in the peril group cautioned, “Not sure this type of AI can be effective when considering human bio individuality,” raising concerns that algorithmic personalization might oversimplify the nuanced realities of patient care. 3. Data Bias and Accuracy Concerns around data bias and accuracy were especially pronounced among Peril respondents. One physician cautioned, “Biases and overdependence on this tool may affect clinical care in some instances,” highlighting fears that flawed algorithms could introduce new risks rather than solve existing ones. Even among those in the Promise group, these issues were not overlooked. A nurse practitioner offered a stark warning: “If the data going in is biased, the AI will make biased decisions. That’s dangerous for vulnerable patients.” A nurse similarly reflected on the fragility of AI’s promise, noting, “Accuracy depends on diverse, clean data. We’re not there yet, and I worry about how this affects diagnoses.” An administrative staff member underscored the need for strict standards, arguing, “We need rigorous validation for AI systems to ensure accuracy before they’re used in real settings.” 4. Efficient and Effective Care From an operational perspective, many Promise-aligned participants identified efficiency and workflow improvement as key areas where TherapiaAI could add value. A member of the clinical staff explained, “AI can help cut down time spent on repetitive tasks and allow us to focus more on actual patient care,” highlighting how the tool could alleviate day-to-day burdens. A nurse added, “I see potential in using AI to streamline diagnostics and make care more efficient across the board,” while an administrative staff member called it “a game changer for workflow; it simplifies complex data and helps us act quickly.” Still, not all respondents were convinced of its operational benefits. One nurse expressed skepticism, stating, “I wonder if it would be as efficient as a human. I am thinking about the calls that go to AI and often AI cannot determine what to do with the call,” emphasizing concerns that automation may fall short in complex or ambiguous situations. These insights suggest that promise respondents viewed TherapiaAI not just as a conceptual innovation, but as a practical solution for overburdened systems, albeit one whose efficiency still raised some doubts. 5. Enhanced Health Information / Education Access Promise group respondents saw strong potential in improving health information access and patient education. A member of the clinical staff pointed out that “AI tools can provide better access to health information, especially for patients who struggle to navigate the system or don’t always understand what their doctors are saying.” A physician observed that “there’s real promise in using AI to help patients learn about their conditions and treatment options in ways that are easy to understand and culturally appropriate.” Another physician highlighted the tool’s capacity to close longstanding gaps: “Access to information is a huge barrier. If AI can help bridge that gap by giving patients real-time insights and education, that’s a win.” Yet even within this enthusiasm, some respondents voiced caution. A nurse remarked, “Suggesting care plan is good AI use, still keeping final decision to provider,” signaling that while AI-generated education and guidance may be helpful, it must not overstep its role or replace clinician judgment. HealthRiskAI Table 6 highlights perceptions of HealthRiskAI, a tool that integrates patient-specific social determinants of health (SDoH) and provides real-time risk assessments. Accuracy of data was significantly more emphasized by low trust respondents as compared to high trust respondents, reflecting concern over the tool’s reliability. Additionally, the need for a human in the loop finding approached significance for low trust respondents, indicating their desire for continued oversight in AI-assisted processes. In the case of HealthRiskAI, accuracy of data emerged as the top concern for low trust respondents (25.3%) compared to high trust respondents (12.5%), underscoring skepticism around the dependability of AI-generated risk assessments. The need for human oversight was also more frequently cited by low trust participants (16.5%) than high trust participants (7.5%), suggesting that this group sees human involvement as essential for ensuring patient safety and proper interpretation of data. In contrast, high trust respondents were more likely to cite efficient and effective care (15.0% vs. 7.6%) and decision making (11.3% vs. 7.6%) as drivers of trust, signaling confidence in AI’s ability to enhance clinical workflows. The relatively even emphasis on privacy and user interface factors, like maintaining data security, privacy, and anonymity (13.9% low trust vs. 8.8% high trust) and the interplay between AI and user (11.4% low trust vs. 8.8% high trust), indicates these are shared considerations across both groups. Table 6 HealthRiskAI: Reasons for Appeal and Concern Total Peril Promise Sig. Accuracy of data 18.9% 31.8% 16.8% 0.094 Need for human in the loop 11.9% 22.7% 10.2% 0.093 Efficient and effective care 11.3% 0.0% 13.1% Maintaining data security, privacy, & anonymity 11.3% 22.7% 9.5% 0.069 Interplay between AI and user 10.1% 13.6% 9.5% Decision making 9.4% 0.0% 10.9% Questioning if AI is needed 8.8% 13.6% 8.0% Automate documentation 5.7% 0.0% 6.6% Time savings 5.0% 4.5% 5.1% Personalized care 2.5% 4.5% 2.2% Care for the marginalized 1.3% 0.0% 1.5% Constant need to adapt to trends 1.3% 0.0% 1.5% Cost and hesitancy to change 1.3% 0.0% 1.5% Changing provider roles 0.6% 0.0% 0.7% Social determinants of health 0.6% 0.0% 0.7% Unequal access to technology/ infrastructure 0.6% 0.0% 0.7% Other 49.7% 31.8% 52.6% 0.071 Total 159 22 137 Caption : Table 6 presents the themes that emerged from a thematic analysis of open-ended survey data. The frequency of the various themes influencing participants’ perceptions about promises and perils of HealthRiskAI are shown for the total and among promise and peril groups. Appeal and concern differences that were significant at p < = 0.05 are noted. 1. Accuracy of data For those in the promise group, accuracy was acknowledged as a potential weakness that could undermine the tool’s clinical value if left unaddressed. A nurse articulated this tension, noting, “While it seems to be a good tool, there is too much room for error and if one patient doesn't get the care that is needed, that is a problem.” Similarly, a nurse practitioner reflected on the fragility of trust in algorithmic predictions, stating, “It is an algorithm that could always get it wrong and someone can fall through the cracks. If it works, it would be amazing!” These responses reflect conditional optimism: the tool holds real promise, but only if its predictions are validated and its risks acknowledged. On the other hand, respondents from the peril group were more explicitly critical. An administrative staff member described HealthRiskAI as “invasive and prone to error,” voicing deep discomfort with the possibility of algorithmic misjudgment, particularly in sensitive health decisions. 2. Need for human in the loop Consistent with other AI tools evaluated in the study, respondents emphasized the importance of human oversight as essential to responsible implementation. Promise-oriented participants were quick to clarify that HealthRiskAI should complement, rather than replace, clinical judgment. A member of the clinical staff pointed out the limitations of data alone, stating, “Sounds good if it gets an accurate read from my patients, but some of them will just tell a health care professional what they think we want to hear. And evaluating that requires human interaction.” A physician expanded on this concern, explaining, “I appreciate the data access and clinical integration, but I am worried about total reliance and missed diagnoses.” Even among supporters, trust in HealthRiskAI was contingent on the tool being used as part of a human-led decision-making process. As one nurse aptly put it, “Not everything is predictable, everyone is unpredictable,” highlighting the inherent variability in patient behavior and the limitations of even the most advanced models to fully capture clinical nuance. 3. Efficient and effective care Promise-aligned respondents were enthusiastic about the tool’s potential to improve efficiency and care coordination. An administrative staff member described HealthRiskAI as a valuable time-management asset, stating, “This would help so much in providing great time management for providers and providing happy experiences for patients across the board.” The tool was also praised for its potential in high-stakes clinical prioritization. A member of the clinical staff believed the technology could be lifesaving, noting, “This Health Risk AI can help save lives & prioritize sickness to help keep someone well.” Another clinical staff member expressed a more general endorsement, saying simply, “It sounds like this will benefit all parties involved.” Still, even among those optimistic about efficiency, caution remained. A nurse in the peril group warned, “There’s room to lose nuance differences but, overall, it would help get care and intervening to people who need it quickly.” These Promise respondents saw HealthRiskAI as a way to streamline clinical workflows and intervene proactively, if the system works as intended. CommCare Table 7 explores high trust and low trust group perceptions of CommCare, an AI tool designed to evaluate patient communication patterns to enhance engagement and satisfaction. Efficient and effective care was significantly more emphasized by high trust respondents, which indicates that individuals with higher trust in AI view the tool's potential to streamline health care processes and improve patient outcomes as a primary benefit. In contrast, low trust respondents did not emphasize this aspect as significant, possibly reflecting their hesitancy to associate AI with operational efficiency. Table 7 CommCare: Reasons for Appeal and Concern, Total Peril Promise Sig. Efficient and effective care 21.4% 8.3% 23.8% 0.089 Questioning if AI is needed 13.6% 37.5% 9.2% 0 Maintaining data security, privacy, & anonymity 7.8% 12.5% 6.9% Personalized care 6.5% 4.2% 6.9% Time savings 6.5% 12.5% 5.4% Automate documentation 4.5% 0.0% 5.4% Interplay between AI and user 3.9% 0.0% 4.6% Unequal access to technology/ infrastructure 3.2% 4.2% 3.1% Changing provider roles 2.6% 0.0% 3.1% Need for human in the loop 2.6% 8.3% 1.5% 0.055 Accuracy of data 1.3% 4.2% 0.8% Constant need to adapt to trends 1.3% 0.0% 1.5% Decision making 1.3% 0.0% 1.5% Telehealth/telemedicine 1.3% 0.0% 1.5% Administrative tasks 0.6% 4.2% 0.0% Ethical AI 0.6% 0.0% 0.8% Social determinants of health 0.6% 0.0% 0.8% Other 63.6% 62.5% 63.8% Total 154 24 130 Caption : Table 7 presents the themes that emerged from a thematic analysis of open-ended survey data. The frequency of the various themes influencing participants’ perceptions about promises and perils of CommCare are shown for the total and among promise and peril groups. Appeal and concern differences that were significant at p < = 0.05 are noted. 1. Questioning if AI is needed Among respondents in the Peril group, there was significant skepticism about whether such a tool was needed at all. A physician questioned its relevance to everyday clinical practice, stating bluntly, “I don’t think this is needed in my practice.” Similarly, a medical technician categorized the tool as offering little value, remarking, “Not beneficial.” Even among those who technically fell into the Promise group, doubt persisted; one nurse dismissed the tool’s utility outright, saying, “This just seems like a waste of time.” These responses reflect deeper uncertainty not just about the tool’s function, but about whether AI is the right solution for communication-based interventions at all. 2. Efficient and effective care Respondents in the Promise group highlighted CommCare’s potential to enhance efficiency and improve care delivery, particularly when viewed through the lens of patient experience. A physician emphasized that the tool could support “improving patient communication and satisfaction,” a key metric in both quality of care and value-based payment models. A researcher described the tool’s functionality as innovative and accessible, calling it “very good and advanced for me.” Clinical staff reinforced this optimistic framing, with one provider suggesting, “This may greatly improve the client experience,” indicating that, when successful, the tool could meaningfully influence patient engagement and comfort, particularly in high-volume or fragmented care environments. Still, this view was not universal. One nurse in the peril group voiced concern that “there are so many other factors that impact non-compliance with treatment and appointments,” adding that it would be “very surprising if this AI could develop an understanding of the unpredictable patterns of mental illness and preferred communications.” This response underscores how perceptions of efficiency can be tempered by doubts about AI’s ability to address the underlying complexities of care. 3. Maintaining data security, privacy, & anonymity Still, even among the more optimistic Promise respondents, privacy concerns loomed large. A physician raised a critical point about surveillance and perception, asking, “Will big brother be watching? Will patients feel like they are being spied upon?” This discomfort, about both real and perceived intrusiveness, was echoed by others in the Peril group. One physician warned of the broader security risks, stating, “Concern; sensitive information is at risk if [the] system becomes compromised.” These concerns suggest that, regardless of a participant’s overall appraisal of the tool, the ethical implications of monitoring communication patterns are not easily dismissed. Discussion This study explored healthcare providers’ perceptions of health AI in safety-net settings through three guiding research questions: ( 1 ) What benefits are currently perceived? ( 2 ) What concerns are raised? and ( 3 ) How are providers thinking about responsibly navigating these concerns to achieve AI’s benefits? Across these domains, responses revealed a complex interplay between optimism for AI’s potential and persistent concerns about equity, trust, and feasibility. While both high- and low-trust respondents recognized possible advantages, such as enhanced efficiency, improved outcomes, and reduced administrative burden, substantial differences emerged in how each group evaluated potential risks and the conditions under which adoption was deemed acceptable. These perceptions were shaped by contextual factors including trust in technology, digital literacy, and the specific needs of diverse patient populations. Perceived Benefits of Health AI (RQ1) Participants identified a range of perceived benefits related to AI's ability to support efficient, timely, and tailored care in safety-net settings. Across both interviews and surveys, respondents frequently cited improvements in workflow as a core advantage, particularly in high-demand, resource-constrained settings. Tools like DocuScribeAI were seen as especially helpful for reducing administrative burden, allowing providers to redirect time and attention to patient interaction. AI-driven decision-support systems, such as TherapiaAI and HealthRiskAI, were praised for enabling earlier risk identification and supporting more personalized treatment strategies, an asset for managing complex or high-risk cases. High-trust respondents were more likely to frame AI as a complementary force within existing clinical systems, reinforcing rather than replacing provider decision-making. They emphasized its utility in speeding up diagnostics, supporting telehealth expansion, and alleviating pressure on overburdened staff. These respondents also highlighted AI’s potential to extend access and continuity of care in communities where infrastructure is strained and staffing is limited. Overall, providers viewed AI not just as a time-saving tool, but as a strategic asset for improving care coordination, reducing burnout, and potentially improving outcomes for underserved patients. Concerns Limiting Adoption (RQ2) Despite these perceived advantages, providers raised serious concerns about the risks and barriers associated with AI adoption. One of the most frequently mentioned challenges was the fragility of trust in safety-net settings, particularly where patients have historically experienced neglect or bias. Biases in AI algorithms, inadequate representation of rural patient populations in training data, and lack of awareness about the unique health care needs of rural communities may exacerbate health inequities ( 19 ).AI's lack of transparency may lead health care providers to mistrust AI-based clinical decision support systems due to unidentified risks, hindering extensive adoption within the health care setting ( 34 ). Many respondents expressed apprehension about the fairness of AI systems, citing risks associated with biased or non-representative training datasets that could reinforce disparities in diagnosis or treatment recommendations. Data security and patient privacy were also prominent concerns, especially for communication-based tools like CommCare. Respondents worried about unauthorized access to sensitive health information, especially in communities already skeptical of digital surveillance or healthcare institutions. Infrastructural limitations presented further complications. Providers referenced outdated EHR systems, poor broadband connectivity, and limited IT support as serious obstacles to implementation. Such challenges are especially acute in rural hospitals and clinics. Financial constraints were another major barrier, particularly for organizations already operating under narrow margins. Without clear evidence of return on investment or funding support, providers questioned whether AI would be a viable option. Low-trust respondents expressed heightened concern over the opacity of AI systems, pointing to a lack of accountability, unclear oversight mechanisms, and fears of eroding the clinician-patient relationship. These respondents stressed that any tool used in clinical care must be subject to rigorous scrutiny, especially if it has the potential to shape patient outcomes. Navigating Concerns to Achieve Responsible Use (RQ3) In addressing how to responsibly navigate these concerns, providers emphasized the importance of preserving human oversight and ensuring AI functions as a support tool rather than an autonomous decision-maker. There was near-universal endorsement of “human-in-the-loop” models, in which clinicians remain at the center of patient care, using AI to enhance rather than replace their judgment. Addressing data limitations and representation issues requires diversifying study populations, incorporating social determinants of health into AI initiatives, and involving rural stakeholders in the development process ( 31 ). Many respondents called for stronger governance structures, including transparent validation protocols, clear vendor accountability, and collaborative design processes that include frontline users. For AI to be adopted in safety-net settings, respondents stressed that tools must be adaptable to local realities. This includes building culturally appropriate patient education materials, ensuring digital literacy resources are available, and developing consent processes that reflect patients’ lived experiences and levels of trust in the system. Alignment and Tensions with Conventional Literature Narratives Provider perspectives in this study align with literature that stresses the importance of equity, transparency, and human-in-the-loop frameworks ( 4 , 20 ). Much like Brewer et al. (2020) and Abramoff et al. (2023), participants emphasized the risk of AI tools exacerbating inequities if they are not designed with representative data and adequate oversight. In particular, concerns around biased algorithms and the lack of rural or underserved population representation in training datasets ( 19 ) were echoed by both low- and high-trust respondents in our study. The need for diverse, community-informed data to inform model development was frequently cited in both the literature and our qualitative responses, especially in the context of hypothetical tools like TherapiaAI and HealthRiskAI. However, while the literature often adopts a tone of caution or even skepticism, emphasizing barriers such as infrastructural inadequacy ( 2 , 12 – 16 ), digital illiteracy ( 7 , 22 ), and data governance challenges ( 4 , 34 ), our findings add nuance by capturing providers’ conditional optimism. Rather than expressing blanket resistance, many providers in our study articulated a willingness to adopt AI tools, especially if those tools are embedded within accountable governance frameworks, are user-centered in design, and include provider oversight. This pragmatism in response suggests a shift from the literature’s focus on why AI fails to reach vulnerable settings to how it might succeed, given the right conditions. For example, prior work by Browning et al. (2020) and Rajkomar et al. (2018) highlights the integration and interoperability limitations in safety-net environments, particularly due to fragmented EHR systems and limited budgets ( 21 , 34 ). Our participants confirmed these constraints, yet many still expressed enthusiasm for tools like DocuScribeAI and HealthRiskAI, noting that such tools could reduce provider burnout and improve efficiency, if they are designed to integrate seamlessly into existing workflows and are supported by vendor-neutral platforms. These frontline insights extend the literature by emphasizing solution-focused strategies, such as tailoring AI for legacy EHR systems, using cloud-based deployments, and supporting training and change management within resource-limited clinics. Additionally, digital literacy and trust remain key concerns in both the literature and our findings. Hollimon et al. (2025) and Perzynski et al. (2017) stress that access to broadband is insufficient if users do not trust or understand digital tools ( 14 , 17 ). Our provider interviews reinforce this view, with multiple respondents pointing to the labor-intensive process of building patient comfort with AI tools and the need for culturally and linguistically tailored onboarding processes. However, high-trust respondents went a step further, suggesting that AI could enhance equity and engagement if deployed with local buy-in and community engagement, which reflects a more empowered and proactive stance than the largely risk-averse posture seen in many systematic reviews ( 2 , 6 , 7 ). Furthermore, while ethical concerns such as bias, transparency, and accountability dominate much of the academic discourse ( 4 , 29 , 34 ), our findings revealed that providers are simultaneously focused on implementation feasibility: issues like time savings, documentation reduction, and easing diagnostic burdens were emphasized, especially for tools like DocuScribeAI and TherapiaAI. This indicates that providers are not just thinking about whether AI is ethical or fair; they are also weighing whether it helps them deliver better care under existing constraints. Finally, while many studies have called for inclusive design and equitable AI development ( 1 , 5 , 31 ), our findings underscore the urgency of implementation mechanisms that reflect these values. Providers advocated for community-engaged governance models, better infrastructure support, and co-design with clinicians, moving beyond theoretical calls for fairness to practical strategies for trust-building and uptake. Their vision challenges the literature to focus not only on design ethics but also on the day-to-day realities of deploying AI in fragile systems with limited staffing, high patient loads, and fragmented data ecosystems. In sum, our study reaffirms many concerns raised in the literature, particularly around data equity, bias, and infrastructure, and it extends the discussion by highlighting that frontline providers will not be passive recipients of AI tools. Instead, they will be active agents negotiating the trade-offs, envisioning benefits, and advocating for deployment models that fit their contexts. Future research and policy efforts would benefit from centering these provider voices to bridge the gap between theory and practice in responsible AI deployment, especially in settings that, historically, have been overlooked and underserved. Limitations This study has several limitations that warrant consideration. First, the sample, while diverse in terms of provider roles and settings, was limited to healthcare professionals in Texas, which may affect the generalizability of findings to safety-net settings in other geographic regions. Second, the trust coding used to segment respondents into “high-trust” and “low-trust” categories was based on interpretive analysis and may not capture the full nuance of individual perspectives. Third, although the mixed-methods design provided complementary insights, survey and interview data were not collected from the same participants, limiting the ability to directly connect responses. Additionally, social desirability bias may have influenced how some participants discussed their patients’ responses to AI or their institutional readiness. Finally, the study focused primarily on provider perspectives and did not incorporate the voices of patients themselves, which are essential for a more complete understanding of AI’s implications in safety-net care. Conclusion As health AI tools continue to permeate clinical settings, their integration into safety-net healthcare systems presents both profound opportunities and critical risks. This study demonstrates that healthcare providers working in underserved environments are not categorically resistant to AI, rather, their perceptions are shaped by a nuanced evaluation of both promise and peril. While many express cautious optimism about AI’s potential to streamline workflows, personalize care, and reduce burnout, concerns remain acute regarding data integrity, ethical oversight, equity, and infrastructural readiness. One of the study’s central findings is the importance of trust, not just in the technology itself, but in the institutions, data, and processes surrounding its development and deployment. High-trust providers tend to view AI as an augmentative force, capable of enhancing clinical decision-making and expanding care access, particularly through tools like DocuScribeAI and TherapiaAI. Low-trust providers, in contrast, are more attuned to systemic risks: the possibility of bias, the fragility of patient privacy, and the lack of resources to meaningfully support implementation. Despite these differences, there is strong alignment across both groups on the need for human-in-the-loop models and the imperative for AI tools to adapt to the realities of local practice environments. Providers emphasized that successful AI deployment depends not only on technical soundness but also on context-sensitive implementation rooted in cultural competence, clear governance, and community engagement. This research contributes to the growing literature by highlighting how real-world practitioners are actively negotiating the trade-offs of health AI, balancing enthusiasm with critique and hope with caution. Rather than seeing safety-net providers as barriers to innovation, this study reveals them as critical architects of AI innovation and responsible deployment. Their insights offer a roadmap for developers, policymakers, and health systems alike. Moving forward, efforts to integrate AI into safety-net care must prioritize co-design with frontline providers, equitable infrastructure investment, and robust accountability mechanisms. Future research should expand to include patient perspectives, longitudinal evaluations of AI implementation, and analyses of policy levers that can enable ethical and inclusive adoption. Ensuring that Health AI fulfills its promise without reinforcing legacy disparities will require sustained collaboration across sectors and a commitment to centering the voices of those most affected by technological change. Declarations Ethics and Consent: This study was reviewed by the Institutional Review Board of the authors’ institution and granted an exempt determination (STUDY00005660). All participants in this study gave consent for their participation, including permission to publish results from the study. Funding: This research was funded by the Episcopal Health Foundation (no grant number). Conflicts of interest/Competing interests: The authors have no conflicts of interest to declare that are relevant to the content of this article. Clinical trial number: Not applicable. Author Contribution CW and MKK designed the study and secured funding, all authors contributed to the data collection and analysis, IP and MKK drafted the manuscript, all authors reviewed the manuscript. References Brewer LC, Fortuna KL, Jones C, Walker R, Hayes SN, Patten CA, et al. Back to the Future: Achieving Health Equity Through Health Informatics and Digital Health. JMIR MHealth UHealth. 2020;8(1):e14512. Marko JGO, Neagu CD, Anand PB. Examining inclusivity: the use of AI and diverse populations in health and social care: a systematic review. BMC Med Inform Decis Mak. 2025;25(1):1–15. Walsh L, Hong SC, Chalakkal RJ, Ogbuehi KC. A Systematic Review of Current Teleophthalmology Services in New Zealand Compared to the Four Comparable Countries of the United Kingdom, Australia, United States of America (USA) and Canada. Vol. 15, CLINICAL OPHTHALMOLOGY. PO BOX 300-008, ALBANY, AUCKLAND 0752, NEW ZEALAND: DOVE MEDICAL PRESS LTD; 2021. p. 4015–27. Rajkomar A, Hardt M, Howell MD, Corrado G, Chin MH. Ensuring Fairness in Machine Learning to Advance Health Equity. Ann Intern Med. 2018;169(12):866. Roy S, Meena T, Lim SJ. Demystifying Supervised Learning in Healthcare 4.0: A New Reality of Transforming Diagnostic Medicine. Vol. 12, DIAGNOSTICS. ST ALBAN-ANLAGE 66, CH-4052 BASEL, SWITZERLAND: MDPI; 2022. Wang JX, Somani S, Chen JH, Murray S, Sarkar U. Health Equity in Artificial Intelligence and Primary Care Research: Protocol for a Scoping Review. JMIR Res Protoc. 2021 Sept 17;10(9):e27799. Perez K, Wisniewski D, Ari A, Lee K, Lieneck C, Ramamonjiarivelo Z. Investigation into Application of AI and Telemedicine in Rural Communities: A Systematic Literature Review. Healthcare. 2025;13(3):324. Al Kuwaiti A, Nazer K, Al-Reedy A, Al-Shehri S, Al-Muhanna A, Subbarayalu AV, et al. A Review of the Role of Artificial Intelligence in Healthcare. J Pers Med. 2023 June 5;13(6):951. Park JI, Bozkurt S, Park JW, Lee S. Evaluation of race/ethnicity-specific survival machine learning models for Hispanic and Black patients with breast cancer. BMJ Health Care Inform [Internet]. 2023 Jan 18 [cited 2025 Mar 5];30(1). Available from: https://informatics.bmj.com/content/30/1/e100666 The Role of Artificial Intelligence (AI) and machine learning in social work practice [Internet]. [cited 2025 June 14]. Available from: https://wjarr.com/content/role-artificial-intelligence-ai-and-machine-learning-social-work-practice Zou J, Schiebinger L. Ensuring that biomedical AI benefits diverse populations. Vol. 67, EBIOMEDICINE. RADARWEG 29, 1043 NX AMSTERDAM, NETHERLANDS: ELSEVIER; 2021. Kohli K, Jain B, Patel TA, Eken HN, Dee EC, Torous J. The digital divide in access to broadband internet and mental healthcare. Nat Ment Health. 2024;2(1):88–95. Madu R. Using Technology To Reduce Barriers In Community Health: An Intersection Between Broadband Access, Ehr & Telehealth Adoption In Federally Qualified Health Centers. Diss Theses Open Access [Internet]. 2020; Available from: https://digitalcommons.library.tmc.edu/uthsph_dissertsopen/127 Perzynski AT, Roach MJ, Shick S, Callahan B, Gunzler D, Cebul R, et al. Patient portals and broadband internet inequality. J Am Med Inform Assoc. 2017 Sept 1;24(5):927–32. Graves JM, Abshire DA, Amiri S, Mackelprang JL. Disparities in Technology and Broadband Internet Access Across Rurality: Implications for Health and Education. Fam Community Health. 2021;44(4):257. Rabbani M. Internet price, speed, and disparity: The case of rural healthcare providers in the United States. Telecommun Policy. 2024;48(2):102674. Hollimon LA, Taylor KV, Fiegenbaum R, Carrasco M, Garchitorena Gomez L, Chung D, et al. Redefining and solving the digital divide and exclusion to improve healthcare: going beyond access to include availability, adequacy, acceptability, and affordability. Front Digit Health [Internet]. 2025 Apr 22 [cited 2025 June 16];7. Available from: https://www.frontiersin.org/journals/digital-health/articles/ 10.3389/fdgth.2025.1508686/full Rhinesmith C, Reisdorf B, Bishop M. The ability to pay for broadband. Commun Res Pract [Internet]. 2019 Apr 3 [cited 2025 June 16]; Available from: https://www.tandfonline.com/doi/full/ 10.1080/22041451.2019.1601491 Abramoff MD, Tarver ME, Loyo-Berrios N, Trujillo S, Char D, Obermeyer Z, et al. Considerations for addressing bias in artificial intelligence for health equity. Vol. 6, NPJ DIGITAL MEDICINE. HEIDELBERGER PLATZ 3, BERLIN, 14197, GERMANY: NATURE PORTFOLIO; 2023. Wylezinski LS, Harris CR, Heiser CN, Gray JD, Spurlock CF. Influence of social determinants of health and county vaccination rates on machine learning models to predict COVID-19 case growth in Tennessee. BMJ Health Care Inform [Internet]. 2021 Sept 27 [cited 2025 Mar 5];28(1). Available from: https://informatics.bmj.com/content/28/1/e100439 Browning T, O’Neill T, Ng Y, Fielding JR, Peshock RM. Special Considerations for Integrating Artificial Intelligence Solutions in Urban Safety-Net Hospitals. J Am Coll Radiol. 2020;17(1):171–4. Chaturvedi U, Chauhan SB, Singh I. The impact of artificial intelligence on remote healthcare: Enhancing patient engagement, connectivity, and overcoming challenges. Intell Pharm [Internet]. 2025 Jan 10 [cited 2025 June 20]; Available from: https://www.sciencedirect.com/science/article/pii/S2949866X24001230 Jabbour S, Fouhey D, Shepard S, Valley TS, Kazerooni EA, Banovic N, et al. Measuring the Impact of AI in the Diagnosis of Hospitalized Patients: A Randomized Clinical Vignette Survey Study. JAMA. 2023;330(23):2275. Braun V, Clarke V. Thematic analysis. In: APA handbook of research methods in psychology, Vol 2: Research designs: Quantitative, qualitative, neuropsychological, and biological. Washington, DC, US: American Psychological Association; 2012. p. 57–71. (APA handbooks in psychology®). Braun V, Clarke V. Is thematic analysis used well in health psychology? A critical review of published research, with recommendations for quality practice and reporting. Health Psychol Rev. 2023;0(0):1–24. Jewett C. Doctors Wrestle With A.I. in Patient Care. N Y Times. 2023;D1–D1. Rose C, Chen JH. Learning from the EHR to implement AI in healthcare. Npj Digit Med. 2024;7(1):1–2. Kitsios F, Kamariotou M, Syngelakis AI, Talias MA. Recent Advances of Artificial Intelligence in Healthcare: A Systematic Literature Review. Vol. 13, APPLIED SCIENCES-BASEL. ST ALBAN-ANLAGE 66, CH-4052 BASEL, SWITZERLAND: MDPI; 2023. Mess SA, Mackey AJ, Yarowsky DE. Artificial Intelligence Scribe and Large Language Model Technology in Healthcare Documentation: Advantages, Limitations, and Recommendations. Plast Reconstr Surg - Glob Open. 2025;13(1):e6450. Parekh ADE, Shaikh OA, Simran, Manan S, Hasibuzzaman MdA. Artificial intelligence (AI) in personalized medicine: AI-generated personalized therapy regimens based on genetic and medical history: short communication. Ann Med Surg. 2023;85(11):5831–3. Uche-Anya E, Anyane-Yeboa A, Berzin TM, Ghassemi M, May FP. Artificial intelligence in gastroenterology and hepatology: how to advance clinical practice while ensuring health equity. Vol. 71, GUT. BRITISH MED ASSOC HOUSE, TAVISTOCK SQUARE, LONDON WC1H 9JR, ENGLAND: BMJ PUBLISHING GROUP; 2022. p. 1909–15. Chen Y, Lehmann CU, Malin B. Digital Information Ecosystems in Modern Care Coordination and Patient Care Pathways and the Challenges and Opportunities for AI Solutions. J Med Internet Res. 2024;26(1):e60258. Youn S, Geismar HN, Pinedo M. Planning and scheduling in healthcare for better care coordination: Current understanding, trending topics, and future opportunities. Prod Oper Manag. 2022;31(12):4407–23. Al Kuwaiti A, Nazer K, Al-Reedy A, Al-Shehri S, Al-Muhanna A, Subbarayalu AV, et al. A Review of the Role of Artificial Intelligence in Healthcare. Vol. 13, JOURNAL OF PERSONALIZED MEDICINE. ST ALBAN-ANLAGE 66, CH-4052 BASEL, SWITZERLAND: MDPI; 2023. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 2 posted You are reading this latest preprint version Show more versions Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7013847","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":617235442,"identity":"4c627a67-86b4-408b-9a80-25ace4b28c8c","order_by":0,"name":"Ishani Purohit","email":"","orcid":"","institution":"Orlando College of Osteopathic Medicine","correspondingAuthor":false,"prefix":"","firstName":"Ishani","middleName":"","lastName":"Purohit","suffix":""},{"id":617235443,"identity":"36348bb9-722b-4f1f-95de-d7eba3f24505","order_by":1,"name":"Matt Kammer-Kerwick","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA6ElEQVRIie3OMQrCQBBA0V0CazOQdkHBEwgRQRSjXkUJ5AwWFiOCNh5AQcgVcoQJAW0W0wbSKGkttLEUoxZardpZ7G8GhnkwjJlMfxpHcoGV8LWRemDxCZIPDIgx+oHExdngS1KbJ9vDUiUVGweifBrF1SpaUQYa0lQen4RpBpJISFJ+PSThdbSECrI/ZeBwFDKauTxk0CxrSZLfyQ4cixXkKvsB2hc9SR+PETjiTtAdIoH4QPL6aqk8kIpPW2rje2EsGu219rHh/rzY9Pp2kEfpaBx3g/n0kB415C2Oz2l9d24ymUwmTTd9hFL/2giinwAAAABJRU5ErkJggg==","orcid":"","institution":"The University of Texas at Austin","correspondingAuthor":true,"prefix":"","firstName":"Matt","middleName":"","lastName":"Kammer-Kerwick","suffix":""},{"id":617235444,"identity":"ada8fd92-00a3-4886-a778-1c87b737dad1","order_by":2,"name":"Emily Spandikow","email":"","orcid":"","institution":"The University of Texas at Austin","correspondingAuthor":false,"prefix":"","firstName":"Emily","middleName":"","lastName":"Spandikow","suffix":""},{"id":617235445,"identity":"bf1bfcba-d813-4700-b07a-0de8cdfc2a96","order_by":3,"name":"Gregory Pogue","email":"","orcid":"","institution":"The University of Texas at Austin","correspondingAuthor":false,"prefix":"","firstName":"Gregory","middleName":"","lastName":"Pogue","suffix":""},{"id":617235446,"identity":"2a79fec2-65ce-4fad-abe6-1ca0b2ab4565","order_by":4,"name":"S. Craig Watkins","email":"","orcid":"","institution":"The University of Texas at Austin","correspondingAuthor":false,"prefix":"","firstName":"S.","middleName":"Craig","lastName":"Watkins","suffix":""}],"badges":[],"createdAt":"2025-06-30 19:53:14","currentVersionCode":2,"declarations":"","doi":"10.21203/rs.3.rs-7013847/v2","doiUrl":"https://doi.org/10.21203/rs.3.rs-7013847/v2","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":106211233,"identity":"8210ab76-52e8-48a1-a5e7-be6ac8086c31","added_by":"auto","created_at":"2026-04-06 07:13:44","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":77869,"visible":true,"origin":"","legend":"\u003cp\u003eAverage Appeal and Concern Rating for Each Vignette\u003c/p\u003e\n\u003cp\u003eCaption: Figure 1 displays the mean scores (plus 95% confidence intervals) for the appeal and concern ratings for the four concepts examined in our survey\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7013847/v2/48d3136bf042bbb1e982971d.jpg"},{"id":108492930,"identity":"093a1882-dadf-4774-9d42-958c9e794c00","added_by":"auto","created_at":"2026-05-05 09:59:00","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1024708,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7013847/v2/f42ec360-b879-4ff6-a0da-02bcd27244a9.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"An Exploration of the Promises and Perils of Responsible Deployment of Health AI for Safety Net Populations as Perceived by Healthcare Providers","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe current study examines the responsible deployment of AI in healthcare settings with a particular focus on underserved, safety net populations. These populations can be rural or urban and are typically characterized by their lack of access to healthcare services, lower-income and resource status, uninsured or under-insured status, and high prevalence of chronic disease. We employ a mixed methods approach to study the perceptions of health care providers relative to the promises of responsible deployment of AI and the potential perils that need to be navigated to achieve that goal. Specifically, we employed a two-study design that involved a set of exploratory interviews among thought leaders and practitioners followed by a survey of practitioners to broaden the perspectives included in our analysis. While the responsible deployment of AI in healthcare settings is important overall, we chose to focus on safety net populations because of some of the unique challenges faced by healthcare providers in serving those populations, including for example, high patient to provider ratios, significant constraints on resources, and the presence of greater health disparities among those community members than in the general population.\u003c/p\u003e \u003cp\u003eThe survey used in this exploratory study included both quantitative and qualitative questioning; our focus in this paper is on the qualitative insights obtained. It is important to note that the survey design included a series of structured vignettes that presented health care providers with specific realistic but fictional health AI technologies. Within these vignettes, healthcare practitioners provided ratings for the degree of appeal and the degree of concern associated with deploying the described technology among safety net community members. The survey also asked healthcare providers to provide a qualitative explanation of how they were thinking about responsible deployment of that technology in the context of those perceived points of appeal and points of concern.\u003c/p\u003e \u003cp\u003eWe performed a literature review that allowed us to focus on specific gaps and used the interviews to develop a more structured line of questioning for the survey. From this first phase we established the following research questions:\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\u003cul\u003e \u003cli\u003e \u003cp\u003eRQ1: What benefits are currently perceived to health AI?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eRQ2: What concerns are currently perceived for health AI?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eRQ3: How are healthcare providers thinking about navigating these concerns in ways that allow the benefits of health AI to be achieved responsibly?\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eWe believe that responses to these questions from the perspectives of safety-net healthcare providers will provide unique insights regarding the design and deployment of health AI in general and in under-resourced settings, specifically.\u003c/p\u003e \u003c/div\u003e"},{"header":"Literature Review","content":"\u003cp\u003eArtificial intelligence has the potential to transform healthcare delivery, particularly in rural and underserved settings where resource limitations and workforce shortages create significant barriers to care. These settings, often characterized by high patient-to-provider ratios, fragmented healthcare infrastructure, and limited access to specialty services, stand to benefit greatly from AI-driven solutions aimed at improving diagnostic accuracy, clinical workflows, and patient management (\u003cspan class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e2\u003c/span\u003e). However, the integration of AI into these environments presents unique technical, financial, and ethical challenges that must be addressed to ensure equitable access to high-quality care.\u003c/p\u003e\u003cp\u003eRural safety-net hospitals serve populations that are disproportionately low-income, uninsured, and burdened with chronic diseases, which can complicate care delivery (\u003cspan class=\"CitationRef\"\u003e3\u003c/span\u003e). AI-enabled technologies such as predictive analytics, natural language processing, and decision-support systems offer opportunities to streamline clinical workflows, optimize resource allocation, and mitigate disparities in healthcare access (\u003cspan class=\"CitationRef\"\u003e4\u003c/span\u003e–\u003cspan class=\"CitationRef\"\u003e6\u003c/span\u003e). Studies have shown that AI-based clinical decision support tools can enhance diagnostic accuracy and reduce cognitive overload for clinicians, which is particularly beneficial in low-resource settings where providers often manage a wide range of complex cases with limited specialist support (\u003cspan class=\"CitationRef\"\u003e7\u003c/span\u003e). Additionally, the use of AI-driven administrative automation, such as auto-populating clinical notes, processing billing, and managing electronic health records, can significantly reduce the time providers spend on non-clinical tasks (\u003cspan class=\"CitationRef\"\u003e8\u003c/span\u003e). This is especially valuable in safety-net settings, where limited staffing and high caseloads often lead to documentation burdens that contribute to provider stress and burnout.\u003c/p\u003e\u003cp\u003eA study incorporating specific AI models for Hispanic and Black women diagnosed with breast cancer shows how specific models can further improve diagnostic and treatment accuracy in underserved and underrepresented populations. These models outperformed general models by better predicting survival outcomes based on their unique health profiles. By tailoring AI models to the unique healthcare needs and disparities of specific demographic groups, these technologies can provide more accurate, individualized care, ultimately improving outcomes and promoting equity (\u003cspan class=\"CitationRef\"\u003e9\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eFor example, tests with Woebot and Crisis Textline chatbots showed a significant reduction in anxiety and symptom mitigation compared to historical values for human intervention (\u003cspan class=\"CitationRef\"\u003e10\u003c/span\u003e). If these benefits can be confirmed, these tools could improve equal access, expand availability in terms of hours per day and user capacity, while offering anonymity and reduced stigma. Populations whose health suffers from the challenges of access and cost while simultaneously being underserved, uninsured, or have difficulty accessing care due to their location are safety net populations. These populations, including rural communities, have reduced access to quality health care and usually rely on publicly funded services because of economic hardship and systemic inequities while experiencing limited access to health care and social support. Although the specific factors may differ, structural barriers hinder their access to essential services and opportunities for improved health outcomes.\u003c/p\u003e\u003cp\u003eDespite these advantages, health IT adoption in rural and underserved communities has historically lagged due to systemic barriers such as interoperability issues, digital infrastructure limitations, and financial constraints (\u003cspan class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e11\u003c/span\u003e). Research consistently finds that one of the major barriers to advanced health-based technological solutions in rural areas is inadequate internet connectivity (\u003cspan class=\"CitationRef\"\u003e12\u003c/span\u003e–\u003cspan class=\"CitationRef\"\u003e16\u003c/span\u003e) And while lower-income households in urban areas have access to the internet primarily via mobile phones, they often encounter struggles to afford consistent access to broadband services (\u003cspan class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e18\u003c/span\u003e). In both instances, the deployment of AI-based health solutions associated with remote patient monitoring or digital biomarkers face notable challenges.\u003c/p\u003e\u003cp\u003eThe implementation of AI in these settings must be approached with a focus on equity, feasibility, and long-term sustainability to avoid exacerbating existing healthcare disparities. As Brewer et al. (2020) highlight, AI models trained on predominantly urban and insured populations may fail to generalize to rural patient populations, leading to bias in diagnostic recommendations and clinical predictions. Addressing this issue requires deliberate efforts to ensure inclusive and representative training datasets that reflect the sociodemographic and epidemiological profiles of rural and underserved communities (\u003cspan class=\"CitationRef\"\u003e19\u003c/span\u003e). A study in Tennessee incorporated social determinants of health (SDOH), such as access to healthcare, transportation, and socioeconomic status, to demonstrate the importance of tailoring AI models to local conditions. Machine learning models that integrated county-level SDOH data showed access to healthcare and vaccination rates were critical in predicting COVID-19 risk and could be applied similarly in rural settings to improve AI predictions and health outcomes. Incorporating community-specific data ensures that AI models are more accurate and reflective of the challenges faced by underserved populations, ultimately promoting equity in care delivery (\u003cspan class=\"CitationRef\"\u003e20\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eAdditionally, safety-net hospitals and rural clinics often lack the financial resources to invest in AI-driven infrastructure. Unlike for-profit institutions, these hospitals rely heavily on Medicaid and fixed tax revenues, which limits their ability to adopt new technologies without external funding or policy incentives (\u003cspan class=\"CitationRef\"\u003e21\u003c/span\u003e). AI solutions for these settings must be cost-effective and easily integrated into existing electronic health records (EHRs) to maximize their impact. Browning et al. (2020) also emphasizes the need for vendor-neutral AI platforms that minimize interoperability challenges, as many safety-net hospitals operate with heterogeneous IT systems due to cost-driven procurement decisions.\u003c/p\u003e\u003cp\u003eFurthermore, patient engagement and digital literacy remain critical factors influencing the success of AI-driven health interventions in rural areas. Studies indicate that racial and ethnic minority populations, who often comprise a significant portion of safety-net hospital patients, use digital health tools at lower rates due to barriers such as lack of trust in AI, poor access to broadband, and limited familiarity with health IT systems (\u003cspan class=\"CitationRef\"\u003e7\u003c/span\u003e). A lack of digital literacy among patients and healthcare providers can limit the effective utilization of AI-driven tools (\u003cspan class=\"CitationRef\"\u003e22\u003c/span\u003e). AI implementation strategies must therefore include culturally tailored patient education and community engagement efforts to ensure that underserved populations can meaningfully interact with these technologies.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eStudy Design Overview\u003c/p\u003e\u003cp\u003eThis study utilized a mixed methods approach to explore healthcare professionals’ perceptions of Health AI (HAI) technologies, particularly in the context of safety-net care. The research design included an extensive review of the literature, a health care provider survey, and a series of in-depth interviews with healthcare practitioners and researchers. These components allowed for the triangulation of insights across conceptual, experiential, and practice-based domains.\u003c/p\u003e\u003cp\u003eThe survey component focused on understanding current awareness and general attitudes toward Health AI, as well as reactions to four hypothetical AI tool deployment scenarios. Open-ended responses captured nuanced perspectives about AI’s utility, feasibility, and perceived risks across different clinical and patient contexts. The interviews were used to deepen understanding of themes identified in the literature and survey by eliciting rich, narrative accounts of participants’ real-world experiences and professional insights. All open-ended responses, including those tied to vignettes, were analyzed qualitatively using a shared coding framework.\u003c/p\u003e\u003cp\u003eStudy 1: Interviews\u003c/p\u003e\u003cp\u003eTo explore provider experiences and perspectives on health AI more deeply, the research team conducted 18 in-depth, semi-structured interviews. Fourteen interviews were conducted with healthcare practitioners, and four with research leaders. Participants were recruited through the research team’s professional network, including through referrals from health system partners. The interviews emphasized themes related to healthcare delivery in safety-net settings, with analytic focus placed on approximately six interviews where participants explicitly discussed rural, underinsured, or medically underserved populations.\u003c/p\u003e\u003cp\u003eInterviews were conducted virtually via Zoom and lasted approximately 30 to 60 minutes. The interview guide included open-ended prompts that addressed a range of topics, such as:\u003c/p\u003e\u003cul\u003e \u003cli\u003e \u003cp\u003eLevels of trust in AI and factors that enhance or erode that trust\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003ePerceived benefits and concerns surrounding AI integration in clinical workflows\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eEthical considerations and bias in AI models\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eInfrastructure and interoperability challenges\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003ePatient-provider communication and engagement, especially in low-resource settings\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe role of AI in advancing or undermining health equity\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e\u003cp\u003eAll interviews were transcribed verbatim and anonymized for analysis. We performed thematic analysis (\u003cspan class=\"CitationRef\"\u003e14\u003c/span\u003e) with multiple coders. A qualitative coding framework consisting of 66 individual codes was developed to capture sentiments expressed across the interviews. These codes were grouped into broader thematic domains representing both the “promise” and “peril” of AI in healthcare. Selective coding was used to highlight particularly illustrative narratives related to safety-net implementation. Quotes from the interviews were later compared and aligned with themes that emerged from the survey data.\u003c/p\u003e\u003cp\u003eStudy 2: Surveys\u003c/p\u003e\u003cp\u003eThe second phase of data collection involved a statewide provider-facing survey designed to capture healthcare professionals’ perceptions of health AI, particularly in the context of rural and underserved populations. The survey was piloted in August 2024 within the research team’s healthcare provider network and subsequently fielded through Dynata, a healthcare panel provider, from September 5 to September 24, 2024. A total of 229 complete responses were collected from healthcare providers across Texas, representing a variety of roles and healthcare settings.\u003c/p\u003e\u003cp\u003eWhile the majority of the survey consisted of quantitative items, several open-ended questions invited participants to elaborate on key topics. These questions were positioned directly after scaled questions and functioned as semi-structured qualitative data, offering context and rationale behind respondents’ ratings. The general open-ended questions included:\u003c/p\u003e\u003cul\u003e \u003cli\u003e \u003cp\u003e“In a sentence or two, please describe what impacts your level of trust in AI technologies to support healthcare delivery.”\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cem\u003e(Linked to question which asked respondents to rate their current level of trust in HAI)\u003c/em\u003e \u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e“In a sentence or two, please explain why you feel this way about your patients’ response to AI to support better health outcomes.”\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cem\u003e(Linked to question which asked how responsive patients would be to HAI)\u003c/em\u003e \u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e“Please list two ways you think AI could specifically benefit rural and other safety-net populations in healthcare.”\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e\u003cp\u003eWe performed thematic analysis (\u003cspan class=\"CitationRef\"\u003e14\u003c/span\u003e) with multiple coders and have focused our analysis of emergent themes as appropriate by level of trust and the balance of appeal vs concern. In addition to these questions, the survey also included four health AI vignettes, using an approach similar to (\u003cspan class=\"CitationRef\"\u003e23\u003c/span\u003e). Each vignette described a hypothetical AI tool for clinical use among safety-net populations. The technologies chosen for the vignettes were designed based on needs and interests expressed in the literature. For each vignette, respondents were asked to rate the tool’s appeal and potential concern on a Likert scale, followed by an open-ended prompt: “In a sentence or two, tell us a little more about the appeal you perceive and any concerns you have.”\u003c/p\u003e\u003cp\u003eSummary descriptions for the four vignettes are:\u003c/p\u003e\u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eDocuScribeAI\u003c/b\u003e: A clinical documentation tool that transcribes and organizes provider-patient interactions in real time using natural language processing and integrates with electronic health records (EHRs).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eTherapiaAI\u003c/b\u003e: A tool for treatment personalization that analyzes data from patient histories, genetics, and current conditions to generate customized therapeutic plans within EHR systems.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eHealthRiskAI\u003c/b\u003e: A predictive analytics platform that incorporates social determinants of health (SDoH) to assess real-time patient risk and optimize care allocation for high-need individuals.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eCommCare\u003c/b\u003e: An AI tool that analyzes patient communication patterns to improve engagement and reduce no-shows in high-volume mental health clinics.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e\u003cp\u003eAcross the eight open-ended questions, including the three general prompts and the four vignette-based responses, nearly 200 qualitative responses were collected for each item. Responses were thematically coded using the same coding framework developed for the interview transcripts. Approximately 30 final codes were applied across the dataset, capturing provider perspectives on trust, ethical concerns, bias, implementation feasibility, cost, health literacy, provider benefit, and patient engagement. These codes were subsequently grouped into thematic domains to support interpretation of patterns across the entire dataset. Representative quotes from the survey are presented in the results section to illustrate key findings.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eStudy 1: Narrative Insights from Thought Leader Interviews\u003c/p\u003e \u003cp\u003eOur analysis of the exploratory interviews in Study 1 produced 5 major themes which we discuss next: Trust in AI (Community \u0026amp; Provider Perspective); Ethical Concerns, Bias, and Transparency; Financial and Integration Constraints in Safety-Net Contexts; Digital Literacy \u0026amp; Patient Engagement; and AI Integration Best Practices.\u003c/p\u003e \u003cp\u003eTrust in AI (Community \u0026amp; Provider Perspective)\u003c/p\u003e \u003cp\u003eTrust in AI is a recurring theme in both the literature and provider narratives, especially in safety-net contexts where patient-provider relationships are often shaped by systemic inequities, historical mistrust, and resource scarcity. For healthcare providers to feel confident in AI tools, they must believe in the reliability, clinical relevance, and transparency of algorithmic decision-making. But building trust goes beyond technical performance; it must also encompass cultural, emotional, and social dimensions, particularly for communities that have been underserved or harmed by healthcare systems in the past.\u003c/p\u003e \u003cp\u003eAs a chief operating officer of digital care platform explains, \u0026ldquo;Folks who are traditionally disadvantaged don't trust a lot of structure because they've felt abused, or they haven't felt that access... they have been turned down more than they have been accepted.\u0026rdquo; In her view, earning trust with these populations requires not just offering solutions, but forging a connection grounded in familiarity and empathy:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eYou\u0026rsquo;ve got to align trust and comfort level with anybody who is using whatever we\u0026rsquo;re suggesting... not just point at people and say, \u0026lsquo;I\u0026rsquo;ve got this done for you.\u0026rsquo; What is that human connection? What is that community and societal connection?\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThis highlights the importance of culturally informed outreach and community-based engagement in building trust, not only in AI tools themselves, but in the systems deploying them.\u003c/p\u003e \u003cp\u003eAn executive leader in community health adds that trust is also shaped by demographic and contextual factors, noting that perceptions of AI may vary based on generational, cultural, and service-line considerations. She observes that \u0026ldquo;place of origin, culture, perspective are factors that are contributing to general trust in the system,\u0026rdquo; especially when working with underserved populations. For example, parents may be more skeptical of AI-supported pediatric care than for themselves, signaling how trust can be context-specific:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eAdults are like, \u0026lsquo;Sure, I\u0026rsquo;ll get on a telemedicine visit\u0026rsquo;... but \u0026lsquo;I want my kid to see the pediatrician, because I think the pediatrician needs to touch my kid. So, I can see that manifesting itself in AI as well.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eFinally, one faith-based mental health leader\u0026rsquo;s insight raises a critical concern about the validation and governance of AI tools, particularly in safety-net environments. For AI to be trusted, users must feel confident that there is oversight and accountability built into the system:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eWho\u0026rsquo;s doing the background check on AI?... It\u0026rsquo;s not a compensation for the lack of providers. We\u0026rsquo;re just trying to find a space where AI can support people in rural communities, because we can\u0026rsquo;t get to them.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eTogether, these perspectives suggest that trust in AI must be earned on multiple fronts: technical reliability, cultural responsiveness, generational comfort, and institutional transparency. In safety-net settings, where providers are often the bridge between systems and communities, any AI implementation must be accompanied by careful attention to the social infrastructure that supports trust.\u003c/p\u003e \u003cp\u003eEthical Concerns, Bias, and Transparency\u003c/p\u003e \u003cp\u003eEthical concerns surrounding AI in healthcare are deeply rooted in questions of fairness, access, and unintended consequences. In safety-net settings, where the stakes are high and resources are limited, these concerns are magnified. Without deliberate efforts to address them, AI risks reinforcing longstanding inequities in health outcomes.\u003c/p\u003e \u003cp\u003eAn executive in rural healthcare leadership expresses a concern shared by many: that AI could exacerbate existing disparities by favoring large, well-funded institutions over smaller, rural providers.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eMy biggest concern is that you're going to have two tiers of systems\u0026hellip; Epic is built for large hospitals, and our smaller rural hospitals don't have access to it\u0026hellip; If these tools\u0026hellip; enhance the efficiency of care and quality of care\u0026hellip; you're going to further widen the quality and care gap that exists today.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eHis remarks highlight the ethical challenge of ensuring equitable access to AI tools, especially when those tools are designed with high-resource settings in mind. This observation illuminates a recurring concern about innovations in healthcare: that they tend to be built for more resource-rich healthcare providers, thus often overlooking the unique needs, challenges, and expertise of resource-poor healthcare providers.\u003c/p\u003e \u003cp\u003eAn expert in health care policy echoes this concern, drawing a parallel to the early days of telehealth. She points out that while large academic centers have the infrastructure to implement new platforms, rural settings often lack both governance and trust.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e\u0026ldquo;In rural settings\u0026hellip; they don't have the governance structure\u0026hellip; That\u0026rsquo;s a barrier because they don\u0026rsquo;t have the resources\u0026hellip; and also trust\u0026hellip; AI reminds me of how telehealth was 15\u0026ndash;20 years ago\u0026hellip; And I think that's where AI is going to be in a few years.\u0026rdquo; Her reflection signals the risk of repeating past mistakes in rolling out health technologies without equity-focused planning.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eA health data innovation leader emphasizes the dual potential of AI, to either narrow or widen disparities, depending on how it is implemented.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e\u0026ldquo;AI has the\u0026hellip; opportunity to close the gap, but if we aren't trying to close the gaps, it will create a bigger gap\u0026hellip; I live in a rural area, and it\u0026rsquo;s about\u0026hellip; real, true access to health care.\u0026rdquo; Her comment reinforces the idea that ethical implementation is not automatic; it requires intentional strategies that prioritize greater attention to the context-specific needs of safety-net populations.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eFinally, an academic innovation researcher calls attention to the often-overlooked ways bias can creep into administrative systems.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e\u0026ldquo;AI in health system administration is probably under-observed\u0026hellip; Hospitals may be unknowingly building biases into their administrative workflows.\u0026rdquo; This observation broadens the conversation about bias beyond clinical algorithms to the operational processes that structure how care is delivered.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eTogether, these perspectives underscore that ethical AI implementation requires more than good intentions. It demands active attention to the ways that bias, access, and transparency shape outcomes and a commitment to designing systems that work for the communities most at risk of being left behind.\u003c/p\u003e \u003cp\u003eFinancial and Integration Constraints in Safety-Net Contexts\u003c/p\u003e \u003cp\u003eThe promise of AI in healthcare often clashes with the practical realities faced by safety-net institutions, where limited budgets, infrastructure gaps, and governance challenges present major barriers to adoption. In these contexts, financial constraints and technological fragmentation make implementation especially difficult.\u003c/p\u003e \u003cp\u003eAn executive in rural healthcare leadership highlights a core financial barrier: without immediate, demonstrable return on investment (ROI), rural hospitals are unlikely to take on the risk of adopting AI tools.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eRural hospitals... would have to have an immediate ROI. A lot of them are cash strapped, as it is... you have to be able to show the savings to get them to buy into it, and then you still have the other barriers, like distrust and misunderstanding and all those other things.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThis participant\u0026rsquo;s comments reflect the pressure rural facilities face to justify every dollar spent, especially to decision-makers on local boards who may lack technical expertise or familiarity with digital health investments.\u003c/p\u003e \u003cp\u003eA faith based mental health leader adds that technological infrastructure remains a major obstacle, particularly in geographically isolated areas with unreliable internet access.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eThose people who are in rural communities... they are also in heavily wooded, non-internet-savvy communities... We provide services in a rural school district... but we still don't have the technology capabilities for them to video into a counselor because the internet is so spotty.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThis leader\u0026rsquo;s example illustrates how integration challenges are not just about software compatibility, but also about physical connectivity and basic digital access.\u003c/p\u003e \u003cp\u003eTogether, these perspectives underscore that financial feasibility and technological readiness must be foundational considerations when designing and deploying AI in safety-net settings. Without affordable solutions and the infrastructure to support them, even the most promising tools will remain out of reach for the communities that could benefit most.\u003c/p\u003e \u003cp\u003eDigital Literacy \u0026amp; Patient Engagement\u003c/p\u003e \u003cp\u003eIn rural and underserved communities, digital literacy and patient engagement are central to the successful adoption of AI tools. While the proliferation of smartphones and internet-enabled devices offers new opportunities to extend care, these tools must align with how people actually use technology, and more importantly, how they feel about using it.\u003c/p\u003e \u003cp\u003eA chief operating officer of digital care platform emphasizes the need to build on existing comfort levels with everyday technologies like mobile phones.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e\u0026ldquo;No matter what else is going on in the world, they have a phone. So how do we seize that comfort level with AI? So how do we adapt to how people use technology?... it's how are we assuring that what we're developing truly will enhance who they are... and enable them to have better health outcomes... not only to care, but to food and to education and to resources.\u0026rdquo; Her framing suggests that AI implementation must start with a deep understanding of how people already engage with technology and how it can be adapted to support broader goals of well-being.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eA health equity executive (D.P) adds that digital literacy is not just about access but also trust and confidence in the systems being offered.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e\u0026ldquo;Our community health workers have to do a lot of one-on-one problem solving to help patients be comfortable with the technology... there's concerns about information security and quality of care. So, I think that's something that we need to work on with communities\u0026hellip; is this a quality and secure alternative for them?\u0026rdquo; She highlights the labor-intensive effort often required to onboard patients into new systems, especially in communities where healthcare technology is still viewed with caution or skepticism.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eTogether, these insights reinforce that engagement with AI in safety-net settings depends not only on access, but on familiarity, trust, and the perceived value of the technology. Tailoring solutions to community habits and building digital confidence are essential steps toward equitable and sustainable AI adoption.\u003c/p\u003e \u003cp\u003eAI Integration Best Practices\u003c/p\u003e \u003cp\u003eSuccessfully implementing AI in safety-net healthcare settings requires more than adopting the latest tools. It demands intentional design, context-aware planning, and structured collaboration between healthcare organizations and technology vendors. Best practices must be rooted in real-world challenges and informed by the operational complexities that providers face every day.\u003c/p\u003e \u003cp\u003eA community health COO emphasizes the regulatory and practical realities of providing care in safety-net systems.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e\u0026ldquo;In a highly regulated industry like healthcare... the challenge is balancing face time with patients while handling documentation and regulatory requirements... those things that are in the world outside of the four walls of our clinic that impacts a patient\u0026rsquo;s ability to engage in their care... trying to mitigate those challenges on a day to day to day basis.\u0026rdquo; Her insight reflects how integration must consider not only the technology itself but also the broader workflow and compliance burdens that shape provider decision-making.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eA chief operating officer of digital care platform focuses on how AI can be used to better connect underserved patients with local services. She sees opportunity in AI\u0026rsquo;s ability to simplify access and personalize support:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e\u0026ldquo;If I\u0026rsquo;m an underserved person... how do we help that person in a respectful way, prioritize those community resources?... we can prove through artificial intelligence that we are actually having a positive social impact on community outreach and community services and healthcare.\u0026rdquo; Her emphasis on respectful guidance and local relevance suggests that integration efforts must be attuned to the everyday realities and needs of the communities they serve. - Worker in Healthcare. It is perspectives like these that are informed by engagement with safety-net communities and familiarity with their needs that make this specific group of practitioners crucial to the design of relevant AI solutions in contexts like these.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eAn executive in rural healthcare leadership adds that governance and vendor collaboration are critical, particularly in under-resourced systems that may lack internal capacity for AI oversight.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e\u0026ldquo;Most of the rural hospitals are ignorant of the capability of the tools... I do think they need some help... maybe having the vendors create a template and having the internal team discuss it... what is missing from this? What other considerations do we need to think about?\u0026rdquo; His comments point to the importance of co-design and shared accountability between technology providers and health systems.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eTogether, these perspectives suggest that successful AI integration must involve thoughtful workflow alignment, community-centered design, and collaborative governance structures. Integration isn\u0026rsquo;t just about embedding AI into clinical systems; it is also about building the trust, infrastructure, and organizational processes that make it work in practice.\u003c/p\u003e \u003cp\u003eStudy 2: Thematic Analysis of Provider Perceptions about AI in Healthcare\u003c/p\u003e \u003cp\u003eAs described under Methodology above, Study 2 included semi-structured lines of questioning about how participants\u0026rsquo; thought about:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eTheir current level of trust in AI technologies to support healthcare delivery\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eTheir perceptions about their patients\u0026rsquo; response to AI to support better health outcomes\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe ways they thought AI could specifically benefit rural and other safety-net populations in healthcare\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eOur analysis of the first 3 topics includes an overall view of the sample plus a stratification by current level of trust (low vs high).\u003c/p\u003e \u003cp\u003eCurrent Level of Trust\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e presents the themes that emerged from a thematic analysis of open-ended survey data (\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e). It shows the frequency of the various themes influencing participants\u0026rsquo; degree of trust in AI technologies for health care delivery among low trust and high trust groups.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eIssues Impacting Trust in AI in Health Care\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTotal Sample\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLow trust\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eHigh trust\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSig\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBuilding trust\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e39.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e50.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e28.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eData bias and accuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e25.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e33.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e17.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.011\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMaintaining data security, privacy, \u0026amp; anonymity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e13.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e14.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEfficient and effective care\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e13.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e23.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAdequate policy, governance, \u0026amp; oversight\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e9.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAdministrative tasks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e13.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.003\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiagnostic capabilities\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e7.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e11.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.011\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNeed for human in the loop\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e7.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAccuracy of data\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e11.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.004\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAutomate documentation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.019\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePersonalized care\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDesire for human touch\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDecision making\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQuestioning if AI is needed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.022\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInterplay between AI and user\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEnhanced health information / education access\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChanging provider roles\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCost and hesitancy to change\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGovernance structure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTime savings\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCare for the marginalized\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConstant need to adapt to trends\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEthical AI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHealth literacy and trust\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInterplay between language, culture, \u0026amp; physicality\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSocial determinants of health\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWorkforce availability, training, \u0026amp; turnover\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOther\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e23.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e16.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e31.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.011\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e211\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e109\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e102\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eCaption\u003c/em\u003e: Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e \u003cem\u003epresents the themes that emerged from a thematic analysis of open-ended survey data. The frequency of the various themes influencing participants\u0026rsquo; degree of trust in AI technologies for health care delivery are shown for the total and among low trust and high trust groups. Trust differences that were significant at p\u0026thinsp;\u0026lt;\u0026thinsp;=\u0026thinsp;0.05 are noted.\u003c/em\u003e\u003c/p\u003e \u003cp\u003eAs shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, the degree of trust in AI among low-trust respondents and high-trust respondents are influenced by different factors. Among low-trust respondents, over half (50.5%) cited building trust as a core issue, followed closely by data bias and accuracy (33.0%) and maintaining data security, privacy, and anonymity (12.8%). These participants expressed deep concern over the fairness, reliability, and safety of AI systems. In contrast, high-trust respondents were more likely to emphasize the practical benefits of AI: 28.4% mentioned building trust, but significantly more cited efficient and effective care (23.5%), diagnostic capabilities (11.8%), and automating documentation (10.8%) as reasons for their confidence in AI. Additionally, high-trust respondents were far more likely to mention the value of adequate governance and accuracy of data compared to their low-trust counterparts.\u003c/p\u003e \u003cp\u003eBoth groups identified building trust as the most frequently mentioned factor behind their current level of trust. (Note the high trust group includes a majority of respondents stating they have a trust level of \u0026ldquo;somewhat\u0026rdquo;.)\u003c/p\u003e \u003cp\u003eHigh trust respondents were more likely to emphasize the importance of (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) data bias and accuracy; (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) adequate policy, governance, and oversight; and (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) administrative tasks. They also value AI's ability to automate documentation, support diagnostic capabilities, and provide efficient and effective care. Maintaining data security, privacy, and anonymity also significantly influences trust among this group. These findings suggest that individuals with high trust focus on AI's ability to streamline processes; enhance governance; and deliver secure, efficient, and effective outcomes.\u003c/p\u003e \u003cp\u003eFor respondents with lower trust levels, the concerns are more targeted, with two key factors emerging as significant - data bias and accuracy are central concerns as well as maintaining data security, privacy, and anonymity - indicating that low trust respondents are particularly sensitive to the fairness and privacy of AI systems. Additionally, ensuring adequate policy, governance, and oversight significantly impacts trust for this group, reflecting fears around the potential misuse or breaches of sensitive information.\u003c/p\u003e \u003cp\u003eIn the course of our analysis of the qualitative data about trust, four themes emerged, which we identify and discuss below.\u003c/p\u003e \u003cp\u003eBuilding Trust\u003c/p\u003e \u003cp\u003eA central theme that emerged across responses was the foundational role of trust when integrating AI into healthcare. Providers consistently expressed skepticism, largely due to the relative novelty of the technology, the absence of long-term outcome data, and the ongoing need for human oversight.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA nurse in the high-trust group described AI as \u0026ldquo;an emerging technology, one that still needs to get the bugs worked out,\u0026rdquo; reflecting early-stage hesitation.\u003c/p\u003e\u003cp\u003eSimilarly, a graduate student healthcare practitioner remarked, \u0026ldquo;There are not enough studies with long-term outcomes to prove AI is safe and effective,\u0026rdquo; pointing to the limited empirical validation of AI tools in clinical practice.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eEven among those open to its potential, there was strong consensus on maintaining human authority in medical decision-making.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eOne nurse firmly stated, \u0026ldquo;AI in no way should diagnose or recommend any medical guidance,\u0026rdquo; emphasizing that clinical decisions should remain with licensed professionals.\u003c/p\u003e\u003cp\u003eA low-trust respondent echoed this need for caution, stating, \u0026ldquo;It is very new and more studies regarding this need to be done,\u0026rdquo; underscoring how trust remains contingent on evidence and transparency, particularly for those less familiar or confident in the technology.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eIn some clinical situations like radiology, for example, AI's ability for diagnostics is the equivalent to or even better than human professionals. Still, human experts are required to confirm diagnosis, communicate the diagnosis to colleagues, and deploy emotional intelligence and context awareness in working with patients to develop a viable treatment plan (\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eData Bias and Accuracy\u003c/p\u003e \u003cp\u003eConcerns about data bias and accuracy were also prominent, particularly among those who expressed conditional trust in AI.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA physician explained, \u0026ldquo;I trust that AI technology can sift through data... but I do not trust the results 100% since there are many examples of either confabulation or incorrect clinical identification,\u0026rdquo; highlighting the tension between data processing capabilities and clinical reliability.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThese concerns were even more pronounced among low-trust respondents.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eOne physician specifically pointed out, \u0026ldquo;The data AI technology is using is not diverse and inclusive,\u0026rdquo; underscoring the risks of algorithmic bias and the need for datasets that reflect diverse patient populations to ensure equitable care.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eAs we develop deeper knowledge of health AI systems and how to deploy them effectively, the definition of diverse patient populations will require additional thought. For example, the deployment of AI in a rural setting will need to be highly sensitive to the context specific features and lived experiences that define life in a particular rural setting. This means that developing algorithmic solutions that are context aware and culturally sensitive will require approaches to governance that are not fully established.\u003c/p\u003e \u003cp\u003eMaintaining Data Security, Privacy, and Anonymity\u003c/p\u003e \u003cp\u003eMaintaining data security, privacy, and anonymity emerged as another significant determinant of trust, with concerns spanning both high- and low-trust groups.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA nurse noted, \u0026ldquo;The recent media about AI compromising people\u0026rsquo;s privacy online makes me a bit wary,\u0026rdquo; capturing a broader apprehension around digital vulnerability.\u003c/p\u003e\u003cp\u003eA physician added, \u0026ldquo;My biggest concern with AI would be the risk of getting hacked,\u0026rdquo; pointing to fears around system breaches and data misuse.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eEven among high-trust participants, the need for formal safeguards remained essential.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eAs one physician emphasized, \u0026ldquo;I need to know that the AI is going to be HIPAA compliant. The platform it is on also impacts whether I feel safe with it.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eTogether, these perspectives signal that trust is inseparable from strong data governance and transparency around privacy protections.\u003c/p\u003e \u003cp\u003eEfficient and Effective Care\u003c/p\u003e \u003cp\u003eHigh-trust respondents more often emphasized the efficiency and effectiveness of AI in improving healthcare delivery.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA nurse described AI as \u0026ldquo;very effective and sufficient in aiding part of our medical system and staff,\u0026rdquo; illustrating its perceived ability to streamline operations.\u003c/p\u003e\u003cp\u003eA physician supported this view, citing emerging research that suggests AI \u0026ldquo;may help throughput,\u0026rdquo; especially in areas like diagnostic imaging.\u003c/p\u003e\u003cp\u003eAnother physician explained how AI could \u0026ldquo;integrate records and abstract (extract?) important information,\u0026rdquo; noting its utility in managing documentation, generating differential diagnoses, and supporting treatment planning.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eHowever, low-trust perspectives reveal concern that the pursuit of efficiency may come at a cost to patient care.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eAs one physician cautioned, \u0026ldquo;Feel like there is going to be a push from higher ups (whether administration or Medicare/Medicaid) to use AI to cut costs and it will create problems in medicine,\u0026rdquo; highlighting that perceived administrative motives may undermine trust in AI\u0026rsquo;s implementation, particularly when efficiency is prioritized over clinical judgment.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ePerceived Benefits of Health AI for Safety Net and Rural Populations\u003c/p\u003e \u003cp\u003eParticipants identified a variety of ways in which AI could positively impact care delivery for underserved and rural communities. Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e summarizes the most frequently mentioned themes across trust groups, offering a comparative view of how low- and high-trust respondents envision AI contributing to safety-net care.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePotential Perceived Benefits of Health AI for Safety Net and Rural Populations\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLow trust\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eHigh trust\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSig.\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEfficient and effective care\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e23.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e22.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e24.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAdministrative tasks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e19.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e17.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e21.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDecision making\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e16.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e19.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e14.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCare for the marginalized\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e11.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiagnostic capabilities\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e11.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e12.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePersonalized care\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e14.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEnhanced health information / education access\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e11.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLack of time, funds, or resources\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTelehealth/telemedicine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e11.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.003\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSocial determinants of health\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAutomate documentation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInterplay between AI and user\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConstant need to adapt to trends\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eData bias and accuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHealth literacy and trust\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIntense and regulated industry\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChanging provider roles\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCost and hesitancy to change\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMaintaining data security, privacy, \u0026amp; anonymity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOther\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e68.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e67.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e69.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e196\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eCaption\u003c/em\u003e: Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e \u003cem\u003epresents the themes that emerged from a thematic analysis of open-ended survey data. The frequency of the various themes influencing participants\u0026rsquo; perceptions about patients\u0026rsquo; response to AI to support better health outcomes shown for the total and among low trust and high trust groups. Trust differences that were significant at p\u0026thinsp;\u0026lt;\u0026thinsp;=\u0026thinsp;0.05 are noted.\u003c/em\u003e\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e depicts how low trust and high trust respondents perceive the potential benefits of AI for rural and safety net services populations. Among all respondents, \u0026ldquo;efficient and effective health care\u0026rdquo; emerged as the leading Health AI benefit. For low trust respondents, \u0026ldquo;decision making,\u0026rdquo; 38%, was the second leading benefit, followed by \u0026ldquo;administrative tasks.\u0026rdquo; For high trust respondents, \u0026ldquo;administrative tasks,\u0026rdquo; 32%, was the second leading benefit, and \u0026ldquo;personalized care\u0026rdquo; was third. Telehealth/telemedicine was significantly more emphasized by high trust respondents (11%) compared to low trust respondents (1%), highlighting that individuals who trust AI technologies may recognize the potential of AI-powered telehealth solutions to bridge gaps in service delivery and enhance care availability for remote or resource-limited populations.\u003c/p\u003e \u003cp\u003eEfficient and Effective Care (Top code for both trust groups)\u003c/p\u003e \u003cp\u003eAcross both high and low trust groups, the most frequently cited benefit of health AI was its potential to deliver more efficient and effective care.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA physician in the high trust group noted that AI \u0026ldquo;has the potential to lead to more comprehensive patient care and reduce wait times,\u0026rdquo; reflecting confidence in its ability to improve throughput and patient satisfaction.\u003c/p\u003e\u003cp\u003eSimilarly, an administrative staff member observed, \u0026ldquo;Because I think that when they see how good it is, they will become a fan,\u0026rdquo; suggesting that positive experiences with AI could drive wider patient acceptance.\u003c/p\u003e\u003cp\u003eA nurse echoed this sentiment, explaining, \u0026ldquo;Patients want care and they want it fast. They want answers and to feel better. AI would assist in all of these things,\u0026rdquo; highlighting how timeliness and responsiveness are central to patient needs and may be supported by AI-driven tools.\u003c/p\u003e\u003cp\u003eBy contrast, a low-trust physician noted, \u0026ldquo;patients would perceive it as a benefit only, if a noticeable care change affects them.,\u0026rdquo; underscoring the conditional nature of acceptance in settings where skepticism remains high.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eAdministrative Tasks (Emphasized more by high trust group)\u003c/p\u003e \u003cp\u003eAdministrative burden reduction was another key theme, particularly among high trust respondents. The use of AI-based automative systems to support administrative tasks has been one of the earliest and most frequent applications in healthcare (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e). Tasks that are routine in nature have been most amenable to automation. Moreover, the significant time expenditure on administrative tasks like clinical notes, billing, and management of electronic health records contributes to high levels of stress and burnout among healthcare practitioners.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA nurse practitioner stated that \u0026ldquo;AI is likely an increasing tool in many services to include healthcare. It may help to streamline care and reduce administrative tasks,\u0026rdquo; suggesting that automation could help free up provider time for direct patient care.\u003c/p\u003e\u003cp\u003eAn administrative staff member similarly emphasized the benefits of convenience and access, explaining, \u0026ldquo;I\u0026rsquo;m sure they will feel better because they will not have to be struggling and they can get the records very very easy through the technology.\u0026rdquo;\u003c/p\u003e\u003cp\u003eAn allied health professional added that AI \u0026ldquo;can create better connections and source for administrative staff,\u0026rdquo; reinforcing the idea that intelligent systems can enhance internal operations and reduce inefficiencies.\u003c/p\u003e\u003cp\u003eOne school-based nurse in the low trust group expressed caution, stating, \u0026ldquo;I'm not exactly sure how AI fits into nursing care in a public-school setting. It's a busy place and I can see the appeal of AI to help with administrative tasks, but privacy concerns outweigh the potential benefit, in my opinion.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThis view illustrates that for some low-trust respondents, the perceived risks, particularly around privacy, may still overshadow the administrative advantages of AI.\u003c/p\u003e \u003cp\u003eDecision Making (Emphasized more by low trust group)\u003c/p\u003e \u003cp\u003eDecision support was cited more frequently by low trust respondents, indicating that while they recognized AI\u0026rsquo;s potential to assist with clinical judgments, they remained cautious.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA member of clinical staff in the low trust group shared that AI could be useful for \u0026ldquo;understanding and summarizing the clinical issues with the client,\u0026rdquo; pointing to its capacity for synthesizing complex data.\u003c/p\u003e\u003cp\u003eHowever, others expressed reservations: a nurse practitioner emphasized the need for more evidence and security, stating, \u0026ldquo;More information [is] needed to improve confidence in AI. Plus, there remains a lack in tight electronic information security.\u0026rdquo;\u003c/p\u003e\u003cp\u003eA physician echoed these concerns, warning that \u0026ldquo;they would need to be informed that using AI comes with risks.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThese responses reveal that while decision-making support is valued, it must be accompanied by safeguards and transparency to be embraced in lower trust environments. By contrast, high trust respondents conveyed greater confidence in AI\u0026rsquo;s role in improving care through better decision-making.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eOne provider stated, \u0026ldquo;AI has the potential to significantly improve the precision and effectiveness of healthcare. By analyzing vast amounts of data, AI can help in early detection of diseases, personalize treatment plans, and provide continuous monitoring and support.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ePersonalized Care (Emphasized more by high trust group)\u003c/p\u003e \u003cp\u003ePersonalized care was more commonly emphasized by high trust respondents, who viewed AI as a way to enhance, not replace, individualized medicine.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA nurse practitioner explained, \u0026ldquo;They would appreciate quicker care but still want personalized care,\u0026rdquo; recognizing the dual importance of speed and customization.\u003c/p\u003e\u003cp\u003eAnother nurse practitioner added that \u0026ldquo;some patients, especially under 40, may better understand and appreciate the personalization,\u0026rdquo; pointing to generational differences in comfort with technology.\u003c/p\u003e\u003cp\u003eEven among low trust respondents, there was some openness to this potential, with an administrative staff member noting, \u0026ldquo;It will create an easier process if things are artificially modified to what actually fits every patient\u0026rsquo;s needs.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThese responses reflect cautious optimism that AI, when properly tailored, can support more patient-centered approaches.\u003c/p\u003e \u003cp\u003eTelehealth/Telemedicine (Significantly more emphasized by high trust group)\u003c/p\u003e \u003cp\u003eTelehealth and telemedicine were significantly more emphasized by high trust respondents, reflecting greater enthusiasm for technology-enabled remote care.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA clinical staff member noted, \u0026ldquo;We use AI in our everyday use now, and in most cases have had a better experience. So why would we not want that for our healthcare?\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eTheir statement suggests normalization of AI in other aspects of life has helped lay the groundwork for acceptance in medical contexts.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA physician similarly observed, \u0026ldquo;With technology today it\u0026rsquo;s more acceptable for patients to try new [tools],\u0026rdquo;- (R_3Ey20tPbO4KkwgZ) underscoring how digital familiarity can support the adoption of AI-powered telehealth solutions.\u003c/p\u003e\u003cp\u003eHowever, caution remained among some low trust providers; one physician commented, \u0026ldquo;Patients are open to being seen sooner and getting the services they need in a timely manner. However, they need to be informed that using AI comes with risks.\u0026rdquo;- (R_5vZXZ0W1BN8eM4y)\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThis remark highlights a recurring theme throughout the study: trust must be earned through transparency, validation, and meaningful patient education.\u003c/p\u003e \u003cp\u003eHealth AI Benefits for Rural and other Safety-Net Populations\u003c/p\u003e \u003cp\u003eTo better understand how healthcare professionals perceive the impact of AI on rural and other safety-net populations, respondents were asked to identify specific ways AI could improve care delivery in these contexts. These insights provide a more grounded view of how AI might be integrated into settings marked by resource constraints and patient vulnerability. Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e summarizes the organizational factors that providers believe will shape the feasibility and sustainability of Health AI implementation in these environments.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eConsiderations for Organization Adoption of Health AI\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLow trust\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eHigh trust\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSig.\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCost and hesitancy to change\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e23.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e38.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e13.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInterplay between AI and user\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e16.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e20.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e13.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMaintaining data security, privacy, \u0026amp; anonymity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e13.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e13.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eData bias and accuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e11.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWorkforce availability, training, \u0026amp; turnover\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e13.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAdequate policy, governance, \u0026amp; oversight\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e7.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e11.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLack of time, funds, or resources\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e7.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e10.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUnequal access to technology/ infrastructure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.036\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChanging provider roles\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBuilding trust\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDesire for human touch\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAdministrative tasks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAutomate documentation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConstant need to adapt to trends\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.005\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEfficient and effective care\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEnhanced health information / education access\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIntense and regulated industry\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQuestioning if AI is needed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChronic health condition prevention and treatment\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDecision making\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiagnostic capabilities\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOther\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e35.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e35.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e35.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eCaption\u003c/em\u003e: Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e \u003cem\u003epresents the themes that emerged from a thematic analysis of open-ended survey data. The frequency of the various themes influencing participants\u0026rsquo; adoption and deployment of health artificial intelligence for health care delivery are shown for the total and among low trust and high trust groups. Trust differences that were significant at p\u0026thinsp;\u0026lt;\u0026thinsp;=\u0026thinsp;0.05 are noted.\u003c/em\u003e\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e highlights additional factors that respondents believe are relevant to their organization's adoption and deployment of Health AI. Cost and hesitancy to change was significantly more emphasized by low trust respondents, reflecting their concerns about financial barriers and organizational resistance to AI integration. On the other hand, unequal access to technology/infrastructure was significantly more emphasized by high trust respondents, indicating that those with greater confidence in AI are more focused on addressing systemic disparities in resource availability to ensure successful implementation. These findings illustrate that addressing financial and cultural resistance is key for low trust groups, while high trust groups prioritize infrastructure equity as a critical enabler of AI adoption.\u003c/p\u003e \u003cp\u003eCost and Hesitancy to change\u003c/p\u003e \u003cp\u003eCost and hesitancy to change emerged as key concerns, particularly among low-trust respondents. Providers in this group expressed doubts about the feasibility of implementing AI given financial constraints and the potential for disruption to patient-provider dynamics.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eOne nurse emphasized that adoption would be more acceptable \u0026ldquo;if it could be implemented with limited intrusion into the provider and patient relationship and did not require a robust electronic record infrastructure or expensive funding to maintain.\u0026rdquo;\u003c/p\u003e\u003cp\u003eA physician echoed this sentiment, acknowledging AI\u0026rsquo;s potential benefits but noting \u0026ldquo;significant front facing costs,\u0026rdquo; which may deter resource-constrained organizations from exploring these tools further.\u003c/p\u003e\u003cp\u003eEven among high-trust respondents, financial burden remained a cautious undercurrent; one physician remarked, \u0026ldquo;I think it can decrease resource needs (with significant front facing costs),\u0026rdquo; highlighting that while optimism exists, concerns about initial investment persist across trust levels.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eInterplay between AI and user\u003c/p\u003e \u003cp\u003eAnother important theme was the interplay between AI and the end user, including both patients and providers. For low-trust respondents, ease of use was critical.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eOne nurse suggested, \u0026ldquo;Make it very basic so patients will feel comfortable with AI,\u0026rdquo; underscoring concerns about accessibility and digital literacy.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eIn contrast, high-trust respondents viewed user familiarity as a strength.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eAn allied health professional observed that many patients already \u0026ldquo;know about AI and support it,\u0026rdquo; suggesting that acceptance may be growing in some communities.\u003c/p\u003e\u003cp\u003eA healthcare IT professional added that AI integration may help with workforce recruitment, stating that \u0026ldquo;AI can improve the provider experience with EMR and help us recruit the newer crop of providers that are learning about AI in school.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThese perspectives indicate that the user experience, both in terms of design simplicity and professional training, will significantly shape adoption trajectories.\u003c/p\u003e \u003cp\u003eData security, privacy, and anonymity\u003c/p\u003e \u003cp\u003eData security, privacy, and anonymity were persistent concerns across both trust groups.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA member of administrative staff in the low-trust group noted that while AI might be helpful for tasks like translation, \u0026ldquo;privacy-wise it\u0026rsquo;s a bit scary.\u0026rdquo;\u003c/p\u003e\u003cp\u003eA clinical staff member, also from a low-trust group, reinforced the importance of this issue: \u0026ldquo;Maintaining data security and privacy will be extremely important for patient trust.\u0026rdquo;\u003c/p\u003e\u003cp\u003eAt the same time, a high-trust clinical provider pointed to AI\u0026rsquo;s potential to advance equity, stating, \u0026ldquo;Possibly the first and/or most neutral approach to healthcare, removing prejudice, racism, sexism, etc... from the equation.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThese comments highlight the tension between AI\u0026rsquo;s transformative potential and the ethical guardrails required to protect patient data and reinforce public trust, particularly in vulnerable communities.\u003c/p\u003e \u003cp\u003eStudy 2: Thematic Analysis of Provider Reactions to Specific Health AI Tools\u003c/p\u003e \u003cp\u003eThe survey included four vignettes about four hypothetical AI tools, designed to assess the appeal and concern for specific Health AI concepts and technologies. The vignettes presented the opportunity to move from abstract to more specific applications of AI in healthcare. See Appendix for full vignettes, which can be summarized as follows:\u003c/p\u003e \u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eDocuScribeAI is an AI tool designed to assist in transcribing and organizing clinical notes during patient interactions.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eTherapiaAI analyzes real-time data from patient histories, genetic information, and current health conditions. It suggests customized therapeutic plans based on clinical guidelines and patient-specific factors.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eHealthRiskAI integrates patient-specific social determinants of health and provides real-time risk assessments.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eCommCare analyzes communication preferences and engagement patterns, offering insights to improve patient-provider interactions.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e \u003cp\u003eDocuScribe draws from AI-powered transcription technologies, particularly large language models (LLMs), to automate clinical documentation and ease administrative burden. These tools generate structured notes from physician\u0026ndash;patient interactions, improving workflow efficiency and reducing burnout, while enabling more patient-centered care. However, risks such as omissions, fabrications, misattributions, and limited contextual understanding, especially around nonverbal cues and speaker differentiation, can compromise accuracy (\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eTherapiaAI is inspired by AI tools in personalized medicine that use genomic, clinical, and lifestyle data to generate individualized treatment plans. These applications enhance diagnostic precision and enable earlier interventions, especially in complex or ambiguous cases, by synthesizing diverse health inputs. Yet, they also raise concerns around biased data, privacy of sensitive genomic information, and clinician challenges in interpreting complex AI outputs due to limited training and infrastructure (\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eHealthRiskAI builds on AI systems that incorporate patient-specific social determinants of health (SDOH) to improve diagnosis and care navigation (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e). These models dynamically integrate factors such as poverty, transportation access, and vaccination rates to produce real-time risk forecasts and guide equitable care coordination. For these systems to fulfill their promise, designers must ensure bias mitigation, interoperability of data, and transparency in how AI-generated risk insights are used in clinical workflows (\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eCommCare was inspired by AI tools that analyze communication preferences and engagement behaviors to improve patient-provider interactions. Using data from EHRs, portals, and care platforms, these tools identify how patients prefer to receive information (e.g., SMS vs. calls) and can flag early signs of disengagement, prompting proactive outreach (\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e). By aligning communication strategies with patient-specific behaviors, these systems enhance adherence, satisfaction, and trust in care relationships (\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAppeal and concern were measured on 5-point scales, as shown next.\u003c/p\u003e \u003cp\u003eAppeal\u003c/p\u003e\n\u003cp\u003e1) No Appeal\u003c/p\u003e\n\n\u003cp\u003e2) A Little Appeal\u003c/p\u003e\n\n\u003cp\u003e3) Some Appeal\u003c/p\u003e\n\n\u003cp\u003e4) Highly Appealing\u003c/p\u003e\n\n\u003cp\u003e5) Extremely Appealing\u003c/p\u003e\n\u003cp\u003eConcern\u003c/p\u003e\n\u003cp\u003e1) No Concern\u003c/p\u003e\n\n\u003cp\u003e2) A Little Concerning\u003c/p\u003e\n\n\u003cp\u003e3) Some Concern\u003c/p\u003e\n\n\u003cp\u003e4) Highly Concerning\u003c/p\u003e\n\n\u003cp\u003e5) Extremely Concerning\u003c/p\u003e\n\u003cp\u003eFigure \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the mean appeal and concern rating for each vignette, including the 95% confidence interval (CI). Across all AI concepts/tools, the average appeal is greater than the average level of concern. The appeal of CommCare is the lowest and is significantly lower than the most appealing concept, DocuScribeAI.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eCaption: Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e displays the mean scores (plus 95% confidence intervals) for the appeal and concern ratings for the four concepts examined in our survey.\u003c/p\u003e \u003cp\u003eParticipants were also asked: \u0026ldquo;In a sentence or two, tell us a little more about the appeal you perceive [for each fictional AI tool] and any concerns you have.\u0026rdquo; Responses were analyzed thematically and coded. Tables\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e through \u003cspan refid=\"Tab7\" class=\"InternalRef\"\u003e7\u003c/span\u003e show these responses for each vignette, with the results stratified by participants as follows:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eThose labeled Promise in the tables and our discussion of our results gave an appeal rating greater than or equal to their concern rating.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThose labeled Peril gave a concern rating greater than their appeal rating.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eOur analysis of the perceptions about the 4 specific, hypothetical Health AI technologies includes an overall view of the sample plus a stratification based on whether perceptions of appeal were greater than perceptions of concern (promise vs. peril).\u003c/p\u003e \u003cp\u003eDocuScribeAI\u003c/p\u003e \u003cp\u003eData from Vignette 1 highlights that, while DocuScribeAI is widely recognized for its ability to automate documentation and save time, skepticism remains about its accuracy and security, and the need to maintain human oversight. The tool's potential to streamline administrative tasks and improve efficiency is promising yet concerns about data reliability and privacy must be addressed to build trust among skeptical stakeholders. Balancing automation with human involvement and ensuring a governance network is in place are critical to maximizing the perceived benefits while mitigating risks. There were no statistically significant differences in findings about this tool among high trust and low trust groups. See Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eData from Vignette 1 highlights that DocuScribeAI is widely perceived as a valuable tool for automating documentation (49.5% high trust, 38.1% low trust) and saving time (32.0% high trust, 32.4% low trust). Respondents also frequently noted the tool\u0026rsquo;s potential for improving accuracy of data (38.1% high trust, 32.4% low trust) and reducing administrative burden. However, some concerns remain, particularly among low trust respondents, regarding the need to maintain human oversight (26.8% high trust, 32.4% low trust), ensure data security and privacy (14.4% high trust, 11.4% low trust), and validate whether AI is appropriate or even needed in clinical workflows (7.2% high trust, 14.3% low trust).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDocuScribeAI: Reasons for Appeal and Concern\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePeril\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePromise\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSig.\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAutomate documentation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e43.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e21.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e47.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.010\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAccuracy of data\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e35.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e46.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e33.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTime savings\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e32.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e14.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e35.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.029\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNeed for human in the loop\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e29.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e46.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e26.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.036\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMaintaining data security, privacy, \u0026amp; anonymity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e13.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e14.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e12.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQuestioning if AI is needed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e21.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e9.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.059\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBuilding trust\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e10.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e9.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEfficient and effective care\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAdequate policy, governance, \u0026amp; oversight\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e10.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInterplay between AI and user\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e10.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInterplay between language, culture, \u0026amp; physicality\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e14.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.029\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAdministrative tasks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChanging provider roles\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWorkforce availability, training, \u0026amp; turnover\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDecision making\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCost and hesitancy to change\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eData bias and accuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConstant need to adapt to trends\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDesire for human touch\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiagnostic capabilities\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePersonalized care\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLack of time, funds, or resources\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTelehealth/telemedicine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUnequal access to technology/ infrastructure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOther\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e199\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e171\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eCaption\u003c/em\u003e: Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e \u003cem\u003epresents the themes that emerged from a thematic analysis of open-ended survey data. The frequency of the various themes influencing participants\u0026rsquo; perceptions about promises and perils of DocuScribeAI are shown for the total and among promise and peril groups. Appeal and concern differences that were significant at p\u0026thinsp;\u0026lt;\u0026thinsp;=\u0026thinsp;0.05 are noted.\u003c/em\u003e\u003c/p\u003e \u003cp\u003eAlthough both groups recognize the potential benefits of increased efficiency (12.4% high trust, 7.6% low trust), the presence of skepticism, especially related to governance and unintended consequences, highlights the importance of transparent oversight frameworks (6.2% high trust, 9.5% low trust).\u003c/p\u003e\n\u003cp\u003e1. Automate Documentation\u003c/p\u003e\n\u003cp\u003eFor respondents in the Promise group, DocuScribeAI was overwhelmingly viewed as a time-saving solution to the documentation burden that plagues many healthcare professionals.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA physician praised its integration and efficiency, explaining that the tool \u0026ldquo;transcribes in real time and [is] integrated with EMR,\u0026rdquo; and noted it \u0026ldquo;should save time if it does not require extensive editing.\u0026rdquo;\u003c/p\u003e\u003cp\u003eEchoing this, a nurse shared enthusiasm for the tool\u0026rsquo;s ability to remove a frustrating task from clinical workflows, stating, \u0026ldquo;That takes medical transcribing off the table and that\u0026rsquo;s wonderful [for] all hospital staff.\u0026rdquo;\u003c/p\u003e\u003cp\u003eThe reduction in burnout was also a recurring theme, with another physician observing, \u0026ldquo;Providers hate spending so much time on the documentation, that is a major cause of their day-to-day stress. Anything which can alleviate it will decrease burnout.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eHowever, among the Peril group, enthusiasm was tempered by skepticism about accuracy and workflow disruption.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eOne nurse questioned the tool\u0026rsquo;s necessity altogether, asking, \u0026ldquo;Charting isn't so overwhelming that AI is needed. How much time would I spend correcting what AI did??\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThis sentiment underscores the concern that rather than saving time, the technology could introduce new inefficiencies if not carefully implemented.\u003c/p\u003e\n\u003cp\u003e2. Time Savings\u003c/p\u003e\n\u003cp\u003eBeyond automating documentation, promise-oriented respondents emphasized time savings as a key appeal.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eOne physician framed the tool as a way to improve efficiency, remarking, \u0026ldquo;It could save massive amounts of time allowing [providers] to see patients more effectively,\u0026rdquo; though they cautioned that this benefit depends on the tool\u0026rsquo;s accuracy.\u003c/p\u003e\u003cp\u003eA nurse succinctly noted, \u0026ldquo;Decreased paper time,\u0026rdquo; while a nurse practitioner elaborated that DocuScribeAI could \u0026ldquo;streamline documentation, allow more time to engage patients, allow providers to see more patients, complete documentation in clinic, [and have] no need to take work home.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThese responses reflected an optimistic view that, if accurate, the tool could support higher-quality patient care while also improving provider work-life balance.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eHowever, not all respondents shared this confidence; one provider in the Peril group expressed skepticism, stating, \u0026ldquo;I'm concerned that it would take as much time to proofread and edit the AI results as it does to just chart myself,\u0026rdquo; highlighting fears that promised time savings might not materialize in practice.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\n\u003cp\u003e3. Accuracy of Data\u003c/p\u003e\n\u003cp\u003eStill, even among those in the Promise group, concerns about accuracy were not absent.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA nurse practitioner admitted some skepticism, saying, \u0026ldquo;I\u0026rsquo;m concerned that it would take as much time to proofread and edit the AI results as it does to just chart myself.\u0026rdquo;\u003c/p\u003e\u003cp\u003eAnother nurse expressed concern over potential clinical risk, cautioning that \u0026ldquo;mistakes are made, words or medications [could be] heard incorrectly, which could cause major issues if not caught in a timely manner.\u0026rdquo;\u003c/p\u003e\u003cp\u003eA physician from the Peril group echoed this hesitation, noting, \u0026ldquo;It sounds very appealing, but I would be worried about the accuracy of the transcription and documentation.\u0026rdquo;\u003c/p\u003e\u003cp\u003eDespite one participant describing it as \u0026ldquo;nice to be able to go back and review those notes and know that they\u0026rsquo;re likely to be very accurate,\u0026rdquo; the Promise group emphasized that speed alone would not justify the use of the tool if accuracy could not be ensured.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\n\u003cp\u003e4. Need for Human in the Loop\u003c/p\u003e\n\u003cp\u003eThe importance of maintaining a human in the loop was central across both Promise and Peril perspectives. For some, this was a manageable requirement.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA nurse in the Promise group raised a critical point about accountability, asking, \u0026ldquo;What errors will there be anli8od who does the weight of the error fall on?\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThis concern becomes more pointed among those in the Peril group.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA nurse stated, \u0026ldquo;I think it would have to be reviewed for accuracy before completely trusting something like this,\u0026rdquo; suggesting that confidence in AI-generated notes remains fragile without reliable safeguards.\u003c/p\u003e\u003cp\u003eA researcher went further, pointing out that \u0026ldquo;you still need to approve the notes. May not even make sense,\u0026rdquo; illustrating a fear that the tool could introduce confusion or workflow delays if its outputs are not clinically usable without intervention.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eTherapiaAI\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e highlights perceptions of TherapiaAI, a tool designed to enhance treatment personalization. While both low trust and high trust groups emphasized the need for human oversight (19.5% low trust,17.4% high trust) and data bias/accuracy (18.4% low trust, 14.0% high trust), their underlying priorities diverged. Low trust respondents were significantly more focused on building trust (16.1%) and data security, privacy, \u0026amp; anonymity (9.2%), underscoring a need for transparency and protection to build confidence in the tool.\u003c/p\u003e \u003cp\u003eIn contrast, high trust respondents placed greater emphasis on personalized care (23.3%) and efficient and effective care (22.1%), suggesting optimism about AI\u0026rsquo;s role in improving patient outcomes. Both groups also cited enhanced health information and education access (13.8% low trust, 12.8% high trust) and time savings (6.9% low trust, 9.3% high trust) as potential benefits, but differences in emphasis reveal that trust level shapes how the benefits of AI are perceived.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eTherapiaAI: Reasons for Appeal and Concern\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePeril\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePromise\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSig.\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNeed for human in the loop\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e18.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e20.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e18.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePersonalized care\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e18.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e20.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e18.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eData bias and accuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e16.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e25.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e14.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEfficient and effective care\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e16.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e18.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.085\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEnhanced health information / education access\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e13.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e14.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBuilding trust\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e16.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e9.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMaintaining data security, privacy, \u0026amp; anonymity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e16.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTime savings\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e8.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDesire for human touch\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e16.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.014\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiagnostic capabilities\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChanging provider roles\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDecision making\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAccuracy of data\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAdministrative tasks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCost and hesitancy to change\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.034\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAutomate documentation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConstant need to adapt to trends\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInterplay between AI and user\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.008\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQuestioning if AI is needed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGovernance structure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAdequate policy, governance, \u0026amp; oversight\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCare for the marginalized\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOther\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e29.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e16.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e31.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e173\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e149\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eCaption\u003c/em\u003e: Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e \u003cem\u003epresents the themes that emerged from a thematic analysis of open-ended survey data. The frequency of the various themes influencing participants\u0026rsquo; perceptions about promises and perils of TherapiaAI are shown for the total and among promise and peril groups. Appeal and concern differences that were significant at p\u0026thinsp;\u0026lt;\u0026thinsp;=\u0026thinsp;0.05 are noted.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e1. Need for Human in the Loop\u003c/p\u003e\n\u003cp\u003eRespondents uniformly expressed a strong commitment to maintaining a \u0026ldquo;human in the loop\u0026rdquo; when deploying TherapiaAI.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA nurse from the promise group emphasized that \u0026ldquo;there should always be a human reviewing or monitoring AI outputs; machines can\u0026rsquo;t replace clinical judgment and compassion,\u0026rdquo; capturing the belief that no matter how advanced the technology becomes, the empathetic, nuanced insight of a clinician remains irreplaceable.\u003c/p\u003e\u003cp\u003eThis perspective was echoed by a physician from the promise group who stated, \u0026ldquo;I am concerned about AI being misaligned with my values and making decisions without provider involvement,\u0026rdquo; while another nurse reinforced, \u0026ldquo;Even with smart tools, oversight is essential.\u003c/p\u003e\u003cp\u003eWe can\u0026rsquo;t just rely on algorithms and hope for the best.\u0026rdquo; One administrative staff member from the peril offered a pointed caution: \u0026ldquo;There\u0026rsquo;s a potential risk that clinical staff might rely too heavily on this tool, potentially neglecting their own clinical and personal judgment.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eTo prevent this, significant safeguards must be established to ensure that the AI complements rather than replaces the valuable insights and decisions made by healthcare professionals.\u0026rdquo;\u003c/p\u003e\n\u003cp\u003e2. Personalized Care\u003c/p\u003e\n\u003cp\u003ePersonalized care emerged as another critical advantage, especially from the promise perspective.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA graduate student healthcare practitioner highlighted the transformative potential of the tool by noting, \u0026ldquo;AI can tailor care to each individual\u0026rsquo;s needs, which is especially important for marginalized populations who are often overlooked.\u0026rdquo;\u003c/p\u003e\u003cp\u003eThis sentiment was shared by a nurse who observed that \u0026ldquo;with AI, patients might finally receive recommendations that truly reflect their history, preferences, and goals.\u0026rdquo;\u003c/p\u003e\u003cp\u003eFurther reinforcing this promise, another nurse commented that \u0026ldquo;personalized care is essential to building trust and better outcomes, and AI can help us get there if used carefully.\u0026rdquo;\u003c/p\u003e\u003cp\u003eStill, not all respondents were fully convinced - one nurse in the peril group cautioned, \u0026ldquo;Not sure this type of AI can be effective when considering human bio individuality,\u0026rdquo; raising concerns that algorithmic personalization might oversimplify the nuanced realities of patient care.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\n\u003cp\u003e3. Data Bias and Accuracy\u003c/p\u003e\n\u003cp\u003eConcerns around data bias and accuracy were especially pronounced among Peril respondents.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eOne physician cautioned, \u0026ldquo;Biases and overdependence on this tool may affect clinical care in some instances,\u0026rdquo; highlighting fears that flawed algorithms could introduce new risks rather than solve existing ones.\u003c/p\u003e\u003cp\u003eEven among those in the Promise group, these issues were not overlooked. A nurse practitioner offered a stark warning: \u0026ldquo;If the data going in is biased, the AI will make biased decisions.\u003c/p\u003e\u003cp\u003eThat\u0026rsquo;s dangerous for vulnerable patients.\u0026rdquo; A nurse similarly reflected on the fragility of AI\u0026rsquo;s promise, noting, \u0026ldquo;Accuracy depends on diverse, clean data. We\u0026rsquo;re not there yet, and I worry about how this affects diagnoses.\u0026rdquo;\u003c/p\u003e\u003cp\u003eAn administrative staff member underscored the need for strict standards, arguing, \u0026ldquo;We need rigorous validation for AI systems to ensure accuracy before they\u0026rsquo;re used in real settings.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\n\u003cp\u003e4. Efficient and Effective Care\u003c/p\u003e\n\u003cp\u003eFrom an operational perspective, many Promise-aligned participants identified efficiency and workflow improvement as key areas where TherapiaAI could add value.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA member of the clinical staff explained, \u0026ldquo;AI can help cut down time spent on repetitive tasks and allow us to focus more on actual patient care,\u0026rdquo; highlighting how the tool could alleviate day-to-day burdens.\u003c/p\u003e\u003cp\u003eA nurse added, \u0026ldquo;I see potential in using AI to streamline diagnostics and make care more efficient across the board,\u0026rdquo; while an administrative staff member called it \u0026ldquo;a game changer for workflow; it simplifies complex data and helps us act quickly.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eStill, not all respondents were convinced of its operational benefits.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eOne nurse expressed skepticism, stating, \u0026ldquo;I wonder if it would be as efficient as a human. I am thinking about the calls that go to AI and often AI cannot determine what to do with the call,\u0026rdquo; emphasizing concerns that automation may fall short in complex or ambiguous situations.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThese insights suggest that promise respondents viewed TherapiaAI not just as a conceptual innovation, but as a practical solution for overburdened systems, albeit one whose efficiency still raised some doubts.\u003c/p\u003e\n\u003cp\u003e5. Enhanced Health Information / Education Access\u003c/p\u003e\n\u003cp\u003ePromise group respondents saw strong potential in improving health information access and patient education. A member of the clinical staff pointed out that \u0026ldquo;AI tools can provide better access to health information, especially for patients who struggle to navigate the system or don\u0026rsquo;t always understand what their doctors are saying.\u0026rdquo;\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA physician observed that \u0026ldquo;there\u0026rsquo;s real promise in using AI to help patients learn about their conditions and treatment options in ways that are easy to understand and culturally appropriate.\u0026rdquo;\u003c/p\u003e\u003cp\u003eAnother physician highlighted the tool\u0026rsquo;s capacity to close longstanding gaps: \u0026ldquo;Access to information is a huge barrier. If AI can help bridge that gap by giving patients real-time insights and education, that\u0026rsquo;s a win.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eYet even within this enthusiasm, some respondents voiced caution.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA nurse remarked, \u0026ldquo;Suggesting care plan is good AI use, still keeping final decision to provider,\u0026rdquo; signaling that while AI-generated education and guidance may be helpful, it must not overstep its role or replace clinician judgment.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eHealthRiskAI\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e highlights perceptions of HealthRiskAI, a tool that integrates patient-specific social determinants of health (SDoH) and provides real-time risk assessments. Accuracy of data was significantly more emphasized by low trust respondents as compared to high trust respondents, reflecting concern over the tool\u0026rsquo;s reliability. Additionally, the need for a human in the loop finding approached significance for low trust respondents, indicating their desire for continued oversight in AI-assisted processes.\u003c/p\u003e \u003cp\u003eIn the case of HealthRiskAI, accuracy of data emerged as the top concern for low trust respondents (25.3%) compared to high trust respondents (12.5%), underscoring skepticism around the dependability of AI-generated risk assessments. The need for human oversight was also more frequently cited by low trust participants (16.5%) than high trust participants (7.5%), suggesting that this group sees human involvement as essential for ensuring patient safety and proper interpretation of data. In contrast, high trust respondents were more likely to cite efficient and effective care (15.0% vs. 7.6%) and decision making (11.3% vs. 7.6%) as drivers of trust, signaling confidence in AI\u0026rsquo;s ability to enhance clinical workflows. The relatively even emphasis on privacy and user interface factors, like maintaining data security, privacy, and anonymity (13.9% low trust vs. 8.8% high trust) and the interplay between AI and user (11.4% low trust vs. 8.8% high trust), indicates these are shared considerations across both groups.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eHealthRiskAI: Reasons for Appeal and Concern\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePeril\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePromise\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSig.\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAccuracy of data\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e18.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e31.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e16.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.094\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNeed for human in the loop\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e22.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.093\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEfficient and effective care\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e13.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMaintaining data security, privacy, \u0026amp; anonymity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e22.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e9.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.069\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInterplay between AI and user\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e13.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e9.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDecision making\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQuestioning if AI is needed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e13.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e8.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAutomate documentation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTime savings\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePersonalized care\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCare for the marginalized\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConstant need to adapt to trends\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCost and hesitancy to change\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChanging provider roles\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSocial determinants of health\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUnequal access to technology/ infrastructure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOther\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e49.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e31.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e52.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.071\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e159\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e137\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eCaption\u003c/em\u003e: Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e \u003cem\u003epresents the themes that emerged from a thematic analysis of open-ended survey data. The frequency of the various themes influencing participants\u0026rsquo; perceptions about promises and perils of HealthRiskAI are shown for the total and among promise and peril groups. Appeal and concern differences that were significant at p\u0026thinsp;\u0026lt;\u0026thinsp;=\u0026thinsp;0.05 are noted.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e1. Accuracy of data\u003c/p\u003e\n\u003cp\u003eFor those in the promise group, accuracy was acknowledged as a potential weakness that could undermine the tool\u0026rsquo;s clinical value if left unaddressed.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA nurse articulated this tension, noting, \u0026ldquo;While it seems to be a good tool, there is too much room for error and if one patient doesn't get the care that is needed, that is a problem.\u0026rdquo;\u003c/p\u003e\u003cp\u003eSimilarly, a nurse practitioner reflected on the fragility of trust in algorithmic predictions, stating, \u0026ldquo;It is an algorithm that could always get it wrong and someone can fall through the cracks. If it works, it would be amazing!\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThese responses reflect conditional optimism: the tool holds real promise, but only if its predictions are validated and its risks acknowledged. On the other hand, respondents from the peril group were more explicitly critical.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eAn administrative staff member described HealthRiskAI as \u0026ldquo;invasive and prone to error,\u0026rdquo; voicing deep discomfort with the possibility of algorithmic misjudgment, particularly in sensitive health decisions.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\n\u003cp\u003e2. Need for human in the loop\u003c/p\u003e\n\u003cp\u003eConsistent with other AI tools evaluated in the study, respondents emphasized the importance of human oversight as essential to responsible implementation. Promise-oriented participants were quick to clarify that HealthRiskAI should complement, rather than replace, clinical judgment.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA member of the clinical staff pointed out the limitations of data alone, stating, \u0026ldquo;Sounds good if it gets an accurate read from my patients, but some of them will just tell a health care professional what they think we want to hear. And evaluating that requires human interaction.\u0026rdquo;\u003c/p\u003e\u003cp\u003eA physician expanded on this concern, explaining, \u0026ldquo;I appreciate the data access and clinical integration, but I am worried about total reliance and missed diagnoses.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eEven among supporters, trust in HealthRiskAI was contingent on the tool being used as part of a human-led decision-making process.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eAs one nurse aptly put it, \u0026ldquo;Not everything is predictable, everyone is unpredictable,\u0026rdquo; highlighting the inherent variability in patient behavior and the limitations of even the most advanced models to fully capture clinical nuance.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\n\u003cp\u003e3. Efficient and effective care\u003c/p\u003e\n\u003cp\u003ePromise-aligned respondents were enthusiastic about the tool\u0026rsquo;s potential to improve efficiency and care coordination.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eAn administrative staff member described HealthRiskAI as a valuable time-management asset, stating, \u0026ldquo;This would help so much in providing great time management for providers and providing happy experiences for patients across the board.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThe tool was also praised for its potential in high-stakes clinical prioritization.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA member of the clinical staff believed the technology could be lifesaving, noting, \u0026ldquo;This Health Risk AI can help save lives \u0026amp; prioritize sickness to help keep someone well.\u0026rdquo;\u003c/p\u003e\u003cp\u003eAnother clinical staff member expressed a more general endorsement, saying simply, \u0026ldquo;It sounds like this will benefit all parties involved.\u0026rdquo;\u003c/p\u003e\u003cp\u003eStill, even among those optimistic about efficiency, caution remained. A nurse in the peril group warned, \u0026ldquo;There\u0026rsquo;s room to lose nuance differences but, overall, it would help get care and intervening to people who need it quickly.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThese Promise respondents saw HealthRiskAI as a way to streamline clinical workflows and intervene proactively, if the system works as intended.\u003c/p\u003e \u003cp\u003eCommCare\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab7\" class=\"InternalRef\"\u003e7\u003c/span\u003e explores high trust and low trust group perceptions of CommCare, an AI tool designed to evaluate patient communication patterns to enhance engagement and satisfaction. Efficient and effective care was significantly more emphasized by high trust respondents, which indicates that individuals with higher trust in AI view the tool's potential to streamline health care processes and improve patient outcomes as a primary benefit. In contrast, low trust respondents did not emphasize this aspect as significant, possibly reflecting their hesitancy to associate AI with operational efficiency.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab7\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 7\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eCommCare: Reasons for Appeal and Concern,\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePeril\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePromise\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSig.\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEfficient and effective care\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e21.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e23.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.089\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQuestioning if AI is needed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e13.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e37.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e9.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMaintaining data security, privacy, \u0026amp; anonymity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e7.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePersonalized care\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTime savings\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAutomate documentation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInterplay between AI and user\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUnequal access to technology/ infrastructure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChanging provider roles\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNeed for human in the loop\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.055\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAccuracy of data\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConstant need to adapt to trends\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDecision making\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTelehealth/telemedicine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAdministrative tasks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEthical AI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSocial determinants of health\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOther\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e63.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e62.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e63.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e154\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e130\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eCaption\u003c/em\u003e: Table\u0026nbsp;\u003cspan refid=\"Tab7\" class=\"InternalRef\"\u003e7\u003c/span\u003e \u003cem\u003epresents the themes that emerged from a thematic analysis of open-ended survey data. The frequency of the various themes influencing participants\u0026rsquo; perceptions about promises and perils of CommCare are shown for the total and among promise and peril groups. Appeal and concern differences that were significant at p\u0026thinsp;\u0026lt;\u0026thinsp;=\u0026thinsp;0.05 are noted.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e1. Questioning if AI is needed\u003c/p\u003e\n\u003cp\u003eAmong respondents in the Peril group, there was significant skepticism about whether such a tool was needed at all.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA physician questioned its relevance to everyday clinical practice, stating bluntly, \u0026ldquo;I don\u0026rsquo;t think this is needed in my practice.\u0026rdquo;\u003c/p\u003e\u003cp\u003eSimilarly, a medical technician categorized the tool as offering little value, remarking, \u0026ldquo;Not beneficial.\u0026rdquo;\u003c/p\u003e\u003cp\u003eEven among those who technically fell into the Promise group, doubt persisted; one nurse dismissed the tool\u0026rsquo;s utility outright, saying, \u0026ldquo;This just seems like a waste of time.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThese responses reflect deeper uncertainty not just about the tool\u0026rsquo;s function, but about whether AI is the right solution for communication-based interventions at all.\u003c/p\u003e\n\u003cp\u003e2. Efficient and effective care\u003c/p\u003e\n\u003cp\u003eRespondents in the Promise group highlighted CommCare\u0026rsquo;s potential to enhance efficiency and improve care delivery, particularly when viewed through the lens of patient experience.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA physician emphasized that the tool could support \u0026ldquo;improving patient communication and satisfaction,\u0026rdquo; a key metric in both quality of care and value-based payment models.\u003c/p\u003e\u003cp\u003eA researcher described the tool\u0026rsquo;s functionality as innovative and accessible, calling it \u0026ldquo;very good and advanced for me.\u0026rdquo;\u003c/p\u003e\u003cp\u003eClinical staff reinforced this optimistic framing, with one provider suggesting, \u0026ldquo;This may greatly improve the client experience,\u0026rdquo; indicating that, when successful, the tool could meaningfully influence patient engagement and comfort, particularly in high-volume or fragmented care environments.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eStill, this view was not universal.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eOne nurse in the peril group voiced concern that \u0026ldquo;there are so many other factors that impact non-compliance with treatment and appointments,\u0026rdquo; adding that it would be \u0026ldquo;very surprising if this AI could develop an understanding of the unpredictable patterns of mental illness and preferred communications.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThis response underscores how perceptions of efficiency can be tempered by doubts about AI\u0026rsquo;s ability to address the underlying complexities of care.\u003c/p\u003e\n\u003cp\u003e3. Maintaining data security, privacy, \u0026 anonymity\u003c/p\u003e\n\u003cp\u003eStill, even among the more optimistic Promise respondents, privacy concerns loomed large.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eA physician raised a critical point about surveillance and perception, asking, \u0026ldquo;Will big brother be watching? Will patients feel like they are being spied upon?\u0026rdquo;\u003c/p\u003e\u003cp\u003eThis discomfort, about both real and perceived intrusiveness, was echoed by others in the Peril group. One physician warned of the broader security risks, stating, \u0026ldquo;Concern; sensitive information is at risk if [the] system becomes compromised.\u0026rdquo;\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThese concerns suggest that, regardless of a participant\u0026rsquo;s overall appraisal of the tool, the ethical implications of monitoring communication patterns are not easily dismissed.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study explored healthcare providers\u0026rsquo; perceptions of health AI in safety-net settings through three guiding research questions: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) What benefits are currently perceived? (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) What concerns are raised? and (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) How are providers thinking about responsibly navigating these concerns to achieve AI\u0026rsquo;s benefits? Across these domains, responses revealed a complex interplay between optimism for AI\u0026rsquo;s potential and persistent concerns about equity, trust, and feasibility. While both high- and low-trust respondents recognized possible advantages, such as enhanced efficiency, improved outcomes, and reduced administrative burden, substantial differences emerged in how each group evaluated potential risks and the conditions under which adoption was deemed acceptable. These perceptions were shaped by contextual factors including trust in technology, digital literacy, and the specific needs of diverse patient populations.\u003c/p\u003e \u003cp\u003ePerceived Benefits of Health AI (RQ1)\u003c/p\u003e \u003cp\u003eParticipants identified a range of perceived benefits related to AI's ability to support efficient, timely, and tailored care in safety-net settings. Across both interviews and surveys, respondents frequently cited improvements in workflow as a core advantage, particularly in high-demand, resource-constrained settings. Tools like DocuScribeAI were seen as especially helpful for reducing administrative burden, allowing providers to redirect time and attention to patient interaction. AI-driven decision-support systems, such as TherapiaAI and HealthRiskAI, were praised for enabling earlier risk identification and supporting more personalized treatment strategies, an asset for managing complex or high-risk cases.\u003c/p\u003e \u003cp\u003eHigh-trust respondents were more likely to frame AI as a complementary force within existing clinical systems, reinforcing rather than replacing provider decision-making. They emphasized its utility in speeding up diagnostics, supporting telehealth expansion, and alleviating pressure on overburdened staff. These respondents also highlighted AI\u0026rsquo;s potential to extend access and continuity of care in communities where infrastructure is strained and staffing is limited. Overall, providers viewed AI not just as a time-saving tool, but as a strategic asset for improving care coordination, reducing burnout, and potentially improving outcomes for underserved patients.\u003c/p\u003e \u003cp\u003eConcerns Limiting Adoption (RQ2)\u003c/p\u003e \u003cp\u003eDespite these perceived advantages, providers raised serious concerns about the risks and barriers associated with AI adoption. One of the most frequently mentioned challenges was the fragility of trust in safety-net settings, particularly where patients have historically experienced neglect or bias. Biases in AI algorithms, inadequate representation of rural patient populations in\u003c/p\u003e \u003cp\u003etraining data, and lack of awareness about the unique health care needs of rural communities\u003c/p\u003e \u003cp\u003emay exacerbate health inequities (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e).AI's lack of transparency may lead health care providers to mistrust AI-based clinical decision support systems due to unidentified risks, hindering extensive adoption within the health care setting (\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e). Many respondents expressed apprehension about the fairness of AI systems, citing risks associated with biased or non-representative training datasets that could reinforce disparities in diagnosis or treatment recommendations.\u003c/p\u003e \u003cp\u003eData security and patient privacy were also prominent concerns, especially for communication-based tools like CommCare. Respondents worried about unauthorized access to sensitive health information, especially in communities already skeptical of digital surveillance or healthcare institutions. Infrastructural limitations presented further complications. Providers referenced outdated EHR systems, poor broadband connectivity, and limited IT support as serious obstacles to implementation. Such challenges are especially acute in rural hospitals and clinics.\u003c/p\u003e \u003cp\u003eFinancial constraints were another major barrier, particularly for organizations already operating under narrow margins. Without clear evidence of return on investment or funding support, providers questioned whether AI would be a viable option. Low-trust respondents expressed heightened concern over the opacity of AI systems, pointing to a lack of accountability, unclear oversight mechanisms, and fears of eroding the clinician-patient relationship. These respondents stressed that any tool used in clinical care must be subject to rigorous scrutiny, especially if it has the potential to shape patient outcomes.\u003c/p\u003e \u003cp\u003eNavigating Concerns to Achieve Responsible Use (RQ3)\u003c/p\u003e \u003cp\u003eIn addressing how to responsibly navigate these concerns, providers emphasized the importance of preserving human oversight and ensuring AI functions as a support tool rather than an autonomous decision-maker. There was near-universal endorsement of \u0026ldquo;human-in-the-loop\u0026rdquo; models, in which clinicians remain at the center of patient care, using AI to enhance rather than replace their judgment.\u003c/p\u003e \u003cp\u003eAddressing data limitations and representation issues requires diversifying study populations, incorporating social determinants of health into AI initiatives, and involving rural stakeholders in the development process (\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e). Many respondents called for stronger governance structures, including transparent validation protocols, clear vendor accountability, and collaborative design processes that include frontline users. For AI to be adopted in safety-net settings, respondents stressed that tools must be adaptable to local realities. This includes building culturally appropriate patient education materials, ensuring digital literacy resources are available, and developing consent processes that reflect patients\u0026rsquo; lived experiences and levels of trust in the system.\u003c/p\u003e \u003cp\u003eAlignment and Tensions with Conventional Literature Narratives\u003c/p\u003e \u003cp\u003eProvider perspectives in this study align with literature that stresses the importance of equity, transparency, and human-in-the-loop frameworks (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e). Much like Brewer et al. (2020) and Abramoff et al. (2023), participants emphasized the risk of AI tools exacerbating inequities if they are not designed with representative data and adequate oversight. In particular, concerns around biased algorithms and the lack of rural or underserved population representation in training datasets (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e) were echoed by both low- and high-trust respondents in our study. The need for diverse, community-informed data to inform model development was frequently cited in both the literature and our qualitative responses, especially in the context of hypothetical tools like TherapiaAI and HealthRiskAI.\u003c/p\u003e \u003cp\u003eHowever, while the literature often adopts a tone of caution or even skepticism, emphasizing barriers such as infrastructural inadequacy (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan additionalcitationids=\"CR13 CR14 CR15\" citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e), digital illiteracy (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e), and data governance challenges (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e), our findings add nuance by capturing providers\u0026rsquo; conditional optimism. Rather than expressing blanket resistance, many providers in our study articulated a willingness to adopt AI tools, especially if those tools are embedded within accountable governance frameworks, are user-centered in design, and include provider oversight. This pragmatism in response suggests a shift from the literature\u0026rsquo;s focus on why AI fails to reach vulnerable settings to how it might succeed, given the right conditions.\u003c/p\u003e \u003cp\u003eFor example, prior work by Browning et al. (2020) and Rajkomar et al. (2018) highlights the integration and interoperability limitations in safety-net environments, particularly due to fragmented EHR systems and limited budgets (\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e). Our participants confirmed these constraints, yet many still expressed enthusiasm for tools like DocuScribeAI and HealthRiskAI, noting that such tools could reduce provider burnout and improve efficiency, if they are designed to integrate seamlessly into existing workflows and are supported by vendor-neutral platforms. These frontline insights extend the literature by emphasizing solution-focused strategies, such as tailoring AI for legacy EHR systems, using cloud-based deployments, and supporting training and change management within resource-limited clinics.\u003c/p\u003e \u003cp\u003eAdditionally, digital literacy and trust remain key concerns in both the literature and our findings. Hollimon et al. (2025) and Perzynski et al. (2017) stress that access to broadband is insufficient if users do not trust or understand digital tools (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e). Our provider interviews reinforce this view, with multiple respondents pointing to the labor-intensive process of building patient comfort with AI tools and the need for culturally and linguistically tailored onboarding processes. However, high-trust respondents went a step further, suggesting that AI could enhance equity and engagement if deployed with local buy-in and community engagement, which reflects a more empowered and proactive stance than the largely risk-averse posture seen in many systematic reviews (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFurthermore, while ethical concerns such as bias, transparency, and accountability dominate much of the academic discourse (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e), our findings revealed that providers are simultaneously focused on implementation feasibility: issues like time savings, documentation reduction, and easing diagnostic burdens were emphasized, especially for tools like DocuScribeAI and TherapiaAI. This indicates that providers are not just thinking about whether AI is ethical or fair; they are also weighing whether it helps them deliver better care under existing constraints.\u003c/p\u003e \u003cp\u003eFinally, while many studies have called for inclusive design and equitable AI development (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e), our findings underscore the urgency of implementation mechanisms that reflect these values. Providers advocated for community-engaged governance models, better infrastructure support, and co-design with clinicians, moving beyond theoretical calls for fairness to practical strategies for trust-building and uptake. Their vision challenges the literature to focus not only on design ethics but also on the day-to-day realities of deploying AI in fragile systems with limited staffing, high patient loads, and fragmented data ecosystems.\u003c/p\u003e \u003cp\u003eIn sum, our study reaffirms many concerns raised in the literature, particularly around data equity, bias, and infrastructure, and it extends the discussion by highlighting that frontline providers will not be passive recipients of AI tools. Instead, they will be active agents negotiating the trade-offs, envisioning benefits, and advocating for deployment models that fit their contexts. Future research and policy efforts would benefit from centering these provider voices to bridge the gap between theory and practice in responsible AI deployment, especially in settings that, historically, have been overlooked and underserved.\u003c/p\u003e \u003cp\u003eLimitations\u003c/p\u003e \u003cp\u003eThis study has several limitations that warrant consideration. First, the sample, while diverse in terms of provider roles and settings, was limited to healthcare professionals in Texas, which may affect the generalizability of findings to safety-net settings in other geographic regions. Second, the trust coding used to segment respondents into \u0026ldquo;high-trust\u0026rdquo; and \u0026ldquo;low-trust\u0026rdquo; categories was based on interpretive analysis and may not capture the full nuance of individual perspectives. Third, although the mixed-methods design provided complementary insights, survey and interview data were not collected from the same participants, limiting the ability to directly connect responses. Additionally, social desirability bias may have influenced how some participants discussed their patients\u0026rsquo; responses to AI or their institutional readiness. Finally, the study focused primarily on provider perspectives and did not incorporate the voices of patients themselves, which are essential for a more complete understanding of AI\u0026rsquo;s implications in safety-net care.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eAs health AI tools continue to permeate clinical settings, their integration into safety-net healthcare systems presents both profound opportunities and critical risks. This study demonstrates that healthcare providers working in underserved environments are not categorically resistant to AI, rather, their perceptions are shaped by a nuanced evaluation of both promise and peril. While many express cautious optimism about AI\u0026rsquo;s potential to streamline workflows, personalize care, and reduce burnout, concerns remain acute regarding data integrity, ethical oversight, equity, and infrastructural readiness.\u003c/p\u003e \u003cp\u003eOne of the study\u0026rsquo;s central findings is the importance of trust, not just in the technology itself, but in the institutions, data, and processes surrounding its development and deployment. High-trust providers tend to view AI as an augmentative force, capable of enhancing clinical decision-making and expanding care access, particularly through tools like DocuScribeAI and TherapiaAI. Low-trust providers, in contrast, are more attuned to systemic risks: the possibility of bias, the fragility of patient privacy, and the lack of resources to meaningfully support implementation. Despite these differences, there is strong alignment across both groups on the need for human-in-the-loop models and the imperative for AI tools to adapt to the realities of local practice environments. Providers emphasized that successful AI deployment depends not only on technical soundness but also on context-sensitive implementation rooted in cultural competence, clear governance, and community engagement.\u003c/p\u003e \u003cp\u003eThis research contributes to the growing literature by highlighting how real-world practitioners are actively negotiating the trade-offs of health AI, balancing enthusiasm with critique and hope with caution. Rather than seeing safety-net providers as barriers to innovation, this study reveals them as critical architects of AI innovation and responsible deployment. Their insights offer a roadmap for developers, policymakers, and health systems alike. Moving forward, efforts to integrate AI into safety-net care must prioritize co-design with frontline providers, equitable infrastructure investment, and robust accountability mechanisms. Future research should expand to include patient perspectives, longitudinal evaluations of AI implementation, and analyses of policy levers that can enable ethical and inclusive adoption. Ensuring that Health AI fulfills its promise without reinforcing legacy disparities will require sustained collaboration across sectors and a commitment to centering the voices of those most affected by technological change.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e Ethics and Consent: This study was reviewed by the Institutional Review Board of the authors\u0026rsquo; institution and granted an exempt determination (STUDY00005660). All participants in this study gave consent for their participation, including permission to publish results from the study.\u003c/p\u003e\u003ch2\u003eFunding:\u003c/h2\u003e \u003cp\u003eThis research was funded by the Episcopal Health Foundation (no grant number).\u003c/p\u003e \u003cp\u003eConflicts of interest/Competing interests: The authors have no conflicts of interest to declare that are relevant to the content of this article.\u003c/p\u003e \u003cp\u003eClinical trial number: Not applicable.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eCW and MKK designed the study and secured funding, all authors contributed to the data collection and analysis, IP and MKK drafted the manuscript, all authors reviewed the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBrewer LC, Fortuna KL, Jones C, Walker R, Hayes SN, Patten CA, et al. Back to the Future: Achieving Health Equity Through Health Informatics and Digital Health. JMIR MHealth UHealth. 2020;8(1):e14512.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMarko JGO, Neagu CD, Anand PB. Examining inclusivity: the use of AI and diverse populations in health and social care: a systematic review. BMC Med Inform Decis Mak. 2025;25(1):1\u0026ndash;15.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWalsh L, Hong SC, Chalakkal RJ, Ogbuehi KC. A Systematic Review of Current Teleophthalmology Services in New Zealand Compared to the Four Comparable Countries of the United Kingdom, Australia, United States of America (USA) and Canada. Vol. 15, CLINICAL OPHTHALMOLOGY. PO BOX 300-008, ALBANY, AUCKLAND 0752, NEW ZEALAND: DOVE MEDICAL PRESS LTD; 2021. p. 4015\u0026ndash;27.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRajkomar A, Hardt M, Howell MD, Corrado G, Chin MH. Ensuring Fairness in Machine Learning to Advance Health Equity. Ann Intern Med. 2018;169(12):866.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoy S, Meena T, Lim SJ. Demystifying Supervised Learning in Healthcare 4.0: A New Reality of Transforming Diagnostic Medicine. Vol. 12, DIAGNOSTICS. ST ALBAN-ANLAGE 66, CH-4052 BASEL, SWITZERLAND: MDPI; 2022.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang JX, Somani S, Chen JH, Murray S, Sarkar U. Health Equity in Artificial Intelligence and Primary Care Research: Protocol for a Scoping Review. JMIR Res Protoc. 2021 Sept 17;10(9):e27799.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePerez K, Wisniewski D, Ari A, Lee K, Lieneck C, Ramamonjiarivelo Z. Investigation into Application of AI and Telemedicine in Rural Communities: A Systematic Literature Review. Healthcare. 2025;13(3):324.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAl Kuwaiti A, Nazer K, Al-Reedy A, Al-Shehri S, Al-Muhanna A, Subbarayalu AV, et al. A Review of the Role of Artificial Intelligence in Healthcare. J Pers Med. 2023 June 5;13(6):951.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePark JI, Bozkurt S, Park JW, Lee S. Evaluation of race/ethnicity-specific survival machine learning models for Hispanic and Black patients with breast cancer. BMJ Health Care Inform [Internet]. 2023 Jan 18 [cited 2025 Mar 5];30(1). Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://informatics.bmj.com/content/30/1/e100666\u003c/span\u003e\u003cspan address=\"https://informatics.bmj.com/content/30/1/e100666\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThe Role of Artificial Intelligence (AI) and machine learning in social work practice [Internet]. [cited 2025 June 14]. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://wjarr.com/content/role-artificial-intelligence-ai-and-machine-learning-social-work-practice\u003c/span\u003e\u003cspan address=\"https://wjarr.com/content/role-artificial-intelligence-ai-and-machine-learning-social-work-practice\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZou J, Schiebinger L. Ensuring that biomedical AI benefits diverse populations. Vol. 67, EBIOMEDICINE. RADARWEG 29, 1043 NX AMSTERDAM, NETHERLANDS: ELSEVIER; 2021.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKohli K, Jain B, Patel TA, Eken HN, Dee EC, Torous J. The digital divide in access to broadband internet and mental healthcare. Nat Ment Health. 2024;2(1):88\u0026ndash;95.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMadu R. Using Technology To Reduce Barriers In Community Health: An Intersection Between Broadband Access, Ehr \u0026amp; Telehealth Adoption In Federally Qualified Health Centers. Diss Theses Open Access [Internet]. 2020; Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://digitalcommons.library.tmc.edu/uthsph_dissertsopen/127\u003c/span\u003e\u003cspan address=\"https://digitalcommons.library.tmc.edu/uthsph_dissertsopen/127\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePerzynski AT, Roach MJ, Shick S, Callahan B, Gunzler D, Cebul R, et al. Patient portals and broadband internet inequality. J Am Med Inform Assoc. 2017 Sept 1;24(5):927\u0026ndash;32.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGraves JM, Abshire DA, Amiri S, Mackelprang JL. Disparities in Technology and Broadband Internet Access Across Rurality: Implications for Health and Education. Fam Community Health. 2021;44(4):257.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRabbani M. Internet price, speed, and disparity: The case of rural healthcare providers in the United States. Telecommun Policy. 2024;48(2):102674.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHollimon LA, Taylor KV, Fiegenbaum R, Carrasco M, Garchitorena Gomez L, Chung D, et al. Redefining and solving the digital divide and exclusion to improve healthcare: going beyond access to include availability, adequacy, acceptability, and affordability. Front Digit Health [Internet]. 2025 Apr 22 [cited 2025 June 16];7. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.frontiersin.org/journals/digital-health/articles/\u003c/span\u003e\u003cspan address=\"https://www.frontiersin.org/journals/digital-health/articles/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fdgth.2025.1508686/full\u003c/span\u003e\u003cspan address=\"10.3389/fdgth.2025.1508686/full\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRhinesmith C, Reisdorf B, Bishop M. The ability to pay for broadband. Commun Res Pract [Internet]. 2019 Apr 3 [cited 2025 June 16]; Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.tandfonline.com/doi/full/\u003c/span\u003e\u003cspan address=\"https://www.tandfonline.com/doi/full/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1080/22041451.2019.1601491\u003c/span\u003e\u003cspan address=\"10.1080/22041451.2019.1601491\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbramoff MD, Tarver ME, Loyo-Berrios N, Trujillo S, Char D, Obermeyer Z, et al. Considerations for addressing bias in artificial intelligence for health equity. Vol. 6, NPJ DIGITAL MEDICINE. HEIDELBERGER PLATZ 3, BERLIN, 14197, GERMANY: NATURE PORTFOLIO; 2023.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWylezinski LS, Harris CR, Heiser CN, Gray JD, Spurlock CF. Influence of social determinants of health and county vaccination rates on machine learning models to predict COVID-19 case growth in Tennessee. BMJ Health Care Inform [Internet]. 2021 Sept 27 [cited 2025 Mar 5];28(1). Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://informatics.bmj.com/content/28/1/e100439\u003c/span\u003e\u003cspan address=\"https://informatics.bmj.com/content/28/1/e100439\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBrowning T, O\u0026rsquo;Neill T, Ng Y, Fielding JR, Peshock RM. Special Considerations for Integrating Artificial Intelligence Solutions in Urban Safety-Net Hospitals. J Am Coll Radiol. 2020;17(1):171\u0026ndash;4.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChaturvedi U, Chauhan SB, Singh I. The impact of artificial intelligence on remote healthcare: Enhancing patient engagement, connectivity, and overcoming challenges. Intell Pharm [Internet]. 2025 Jan 10 [cited 2025 June 20]; Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.sciencedirect.com/science/article/pii/S2949866X24001230\u003c/span\u003e\u003cspan address=\"https://www.sciencedirect.com/science/article/pii/S2949866X24001230\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJabbour S, Fouhey D, Shepard S, Valley TS, Kazerooni EA, Banovic N, et al. Measuring the Impact of AI in the Diagnosis of Hospitalized Patients: A Randomized Clinical Vignette Survey Study. JAMA. 2023;330(23):2275.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBraun V, Clarke V. Thematic analysis. In: APA handbook of research methods in psychology, Vol 2: Research designs: Quantitative, qualitative, neuropsychological, and biological. Washington, DC, US: American Psychological Association; 2012. p. 57\u0026ndash;71. (APA handbooks in psychology\u0026reg;).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBraun V, Clarke V. Is thematic analysis used well in health psychology? A critical review of published research, with recommendations for quality practice and reporting. Health Psychol Rev. 2023;0(0):1\u0026ndash;24.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJewett C. Doctors Wrestle With A.I. in Patient Care. N Y Times. 2023;D1\u0026ndash;D1.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRose C, Chen JH. Learning from the EHR to implement AI in healthcare. Npj Digit Med. 2024;7(1):1\u0026ndash;2.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKitsios F, Kamariotou M, Syngelakis AI, Talias MA. Recent Advances of Artificial Intelligence in Healthcare: A Systematic Literature Review. Vol. 13, APPLIED SCIENCES-BASEL. ST ALBAN-ANLAGE 66, CH-4052 BASEL, SWITZERLAND: MDPI; 2023.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMess SA, Mackey AJ, Yarowsky DE. Artificial Intelligence Scribe and Large Language Model Technology in Healthcare Documentation: Advantages, Limitations, and Recommendations. Plast Reconstr Surg - Glob Open. 2025;13(1):e6450.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eParekh ADE, Shaikh OA, Simran, Manan S, Hasibuzzaman MdA. Artificial intelligence (AI) in personalized medicine: AI-generated personalized therapy regimens based on genetic and medical history: short communication. Ann Med Surg. 2023;85(11):5831\u0026ndash;3.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eUche-Anya E, Anyane-Yeboa A, Berzin TM, Ghassemi M, May FP. Artificial intelligence in gastroenterology and hepatology: how to advance clinical practice while ensuring health equity. Vol. 71, GUT. BRITISH MED ASSOC HOUSE, TAVISTOCK SQUARE, LONDON WC1H 9JR, ENGLAND: BMJ PUBLISHING GROUP; 2022. p. 1909\u0026ndash;15.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen Y, Lehmann CU, Malin B. Digital Information Ecosystems in Modern Care Coordination and Patient Care Pathways and the Challenges and Opportunities for AI Solutions. J Med Internet Res. 2024;26(1):e60258.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYoun S, Geismar HN, Pinedo M. Planning and scheduling in healthcare for better care coordination: Current understanding, trending topics, and future opportunities. Prod Oper Manag. 2022;31(12):4407\u0026ndash;23.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAl Kuwaiti A, Nazer K, Al-Reedy A, Al-Shehri S, Al-Muhanna A, Subbarayalu AV, et al. A Review of the Role of Artificial Intelligence in Healthcare. Vol. 13, JOURNAL OF PERSONALIZED MEDICINE. ST ALBAN-ANLAGE 66, CH-4052 BASEL, SWITZERLAND: MDPI; 2023.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Responsible AI, Health Equity, Provider Perceptions, Safety Net Populations, Mixed Methods, Vignette Analysis","lastPublishedDoi":"10.21203/rs.3.rs-7013847/v2","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7013847/v2","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe current study examines the responsible deployment of AI in healthcare settings with a particular focus on underserved, safety net populations. We employ a mixed methods approach to study the perceptions of health care providers relative to the promises of responsible deployment of AI and the potential perils that need to be navigated to achieve that goal. As health AI tools continue to enter clinical spaces, understanding how they are perceived by providers in safety-net environments is essential for equitable implementation. This study revealed that while there is cautious optimism among healthcare professionals, particularly regarding improvements in workflow, personalization, and efficiency, significant concerns remain around data integrity, trust, and infrastructural readiness. High-trust providers viewed AI as a valuable support system, whereas low-trust providers raised critical questions about governance, privacy, and the risk of exacerbating existing inequities. The findings emphasize the importance of human-in-the-loop models, localized implementation strategies, and community-informed design to ensure that the promise of AI does not bypass the populations it seeks to serve. Moving forward, engaging providers in policy design, tool development, and implementation processes will be crucial for realizing the equitable integration of AI in healthcare.\u003c/p\u003e","manuscriptTitle":"An Exploration of the Promises and Perils of Responsible Deployment of Health AI for Safety Net Populations as Perceived by Healthcare Providers","msid":"","msnumber":"","nonDraftVersions":[{"code":2,"date":"2026-04-06 07:13:40","doi":"10.21203/rs.3.rs-7013847/v2","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}},{"code":1,"date":"2025-09-04 15:54:27","doi":"10.21203/rs.3.rs-7013847/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"fd160b38-fee1-413e-9df6-aae0c6f3a198","owner":[],"postedDate":"April 6th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Rejected","date":"2026-05-03T20:29:13+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":65688060,"name":"Scientific community and society/Business and industry"},{"id":65688061,"name":"Health sciences/Health care"},{"id":65688062,"name":"Scientific community and society/Scientific community"}],"tags":[],"updatedAt":"2026-05-03T20:40:04+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-06 07:13:40","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v2","identity":"rs-7013847","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7013847","identity":"rs-7013847","version":["v2"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}