From experimentation to adoption: a normative ethical analysis of large language models in health care

From experimentation to adoption: a normative ethical analysis of large language models in health care | 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 Research Article From experimentation to adoption: a normative ethical analysis of large language models in health care Xiongwen Yang, Yuanwei Liang, Yijiang Liu, Di Liu, Lin Yang, Bo Zhang, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8553262/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Large language models (LLMs) are increasingly embedded in routine health-care communication, extending beyond experimental evaluation to real-world use by clinicians, patients, and caregivers. This transition from experimentation to adoption represents a normatively consequential ethical shift. While early research has focused primarily on technical performance, emerging ethical challenges arise less from model capabilities than from how LLMs become integrated into clinical workflows, institutional arrangements, and relationships of care. Methods We conducted a normative ethical analysis informed by empirical evidence from recent real-world studies of LLM adoption in health-care settings. Drawing on adoption-oriented empirical research, we examined how the transition from experimental use to routine reliance reshapes ethical conditions related to trust, responsibility allocation, and equity across different user groups. Results The analysis identifies a systematic shift in ethical risk from model-level concerns—such as accuracy, validation, and bias mitigation—toward system-level dynamics that emerge during routine adoption. Adoption alters how medical information is mediated, redistributes responsibility among clinicians, institutions, and patients, and exposes asymmetrical vulnerability linked to digital literacy, educational background, and regional context. Rather than uniformly democratizing access to medical knowledge, routine LLM use may function as a capability amplifier, magnifying existing inequities in the absence of institutional safeguards. Conclusions Ethical challenges arising from routine LLM adoption in health care cannot be adequately addressed through model-level or pre-deployment ethics alone. Addressing these challenges requires adoption-sensitive governance approaches that recognize reliance as an ongoing ethical process. Institutional accountability, role-sensitive design, and continuous oversight are ethically necessary to ensure that LLM integration enhances health-care communication and access without eroding trust or exacerbating existing disparities. Large language models Medical ethics Ethical governance Health-care communication Technology adoption Equity and trust Responsibility Figures Figure 1 Introduction Large language models (LLMs) have rapidly moved from experimental tools to increasingly visible actors within health-care information ecosystems. Early research has primarily examined whether these systems can perform clinically relevant tasks—such as summarizing medical reports, explaining technical terminology, or supporting decision-making—with acceptable accuracy and efficiency[ 1 – 6 ]. Such work has been essential in establishing proof of concept and translational potential. However, as LLMs are increasingly adopted by clinicians, patients, and caregivers in real-world settings[ 7 – 12 ], a distinct ethical question arises that cannot be answered by performance evaluation alone: what changes ethically when LLMs become objects of routine reliance in everyday care, rather than instruments tested under experimental supervision? We argue that the shift from experimentation to adoption marks an adoption-phase ethics problem: ethical risk increasingly emerges not from isolated model outputs, but from the ways LLMs are embedded into clinical relationships, institutional workflows, and expectations of care. During experimental use, ethical concerns are commonly framed in technical terms—including model performance, validation, and bias mitigation[ 13 ]. In contrast, routine adoption redistributes epistemic authority over medical information, reshapes how trust is negotiated between clinicians and patients, and complicates established lines of professional responsibility[ 9 – 12 ]. These changes are not merely conceptual. Clinicians remain legally and ethically accountable for clinical decisions, yet patients and caregivers may increasingly act on LLM-mediated explanations when interpreting diagnoses, prognoses, and treatment options[ 9 – 12 ]. This configuration creates an ethics of reliance problem: reliance can grow faster than the capacity of users and institutions to interpret uncertainty, contest outputs, or secure recourse when harm occurs. Emerging evidence from adoption-oriented studies reinforces that sustained uptake of LLMs in health care is driven less by algorithmic sophistication than by socio-technical and ethical conditions[ 9 – 12 , 14 ]. Trust, perceived usefulness, digital literacy, privacy assurance, and institutional support consistently shape whether and how LLMs are used[ 9 – 12 , 15 ]. Importantly, trust here is not reducible to confidence in accuracy; it reflects expectations about who stands behind LLM-mediated information, how limitations are signaled, and what protections exist when expectations are violated[ 16 ]. Where such governance signals are weak, adoption may stall—or proceed in informal, uneven, and potentially unsafe ways—particularly among individuals with lower digital literacy or limited access to institutional guidance, raising concerns about inequity and exclusion. These dynamics expose limits in prevailing approaches to AI ethics in medicine, which often assume that improving model performance, explainability, or transparency is sufficient to manage ethical risk[ 17 , 18 ]. As LLMs move into routine clinical and informational use, ethical hazards increasingly arise from institutional embedding rather than from technical properties alone. Transparency without accountable workflows can still produce misplaced trust; explainability without role-appropriate safeguards can still shift burdens onto patients and front-line clinicians. Ethical attention must therefore expand beyond pre-deployment evaluation toward system-level governance questions about trust, responsibility allocation, and equitable capability to benefit. In this Article, adoption is understood not merely as technical deployment, but as the sustained integration of LLMs into everyday clinical and informational practices through which patterns of reliance, trust, and responsibility become socially and institutionally stabilized[ 9 – 12 ]. Using a normative ethical analysis informed by empirical adoption evidence, we examine three interrelated domains—trust, responsibility, and equity—to clarify how routine LLM use can (i) reconfigure institutional and interpersonal trust, (ii) produce responsibility and accountability gaps, and (iii) amplify existing disparities in the capacity to interpret and benefit from AI-mediated information. We focus on health-care communication and non-autonomous decision support rather than fully autonomous diagnosis or treatment recommendation. We conclude by outlining governance implications for ethically robust integration of LLMs into health-care communication. Methods Study design and ethical methodology This study employs a normative ethical analysis informed by empirical evidence, focusing on ethical challenges that arise during the adoption phase of LLMs in health care. The aim is not to evaluate technical performance or quantify adoption outcomes, but to identify ethically salient patterns that emerge when LLMs move from experimental evaluation to routine use in clinical communication and informational contexts. Consistent with established approaches in medical ethics, empirical findings are used as contextual inputs to normative reasoning rather than as sources of causal inference. Empirical evidence is mobilized to illuminate how trust relationships, responsibility allocation, and equity conditions are reconfigured through sustained reliance on LLM-mediated information. Identification and selection of empirical evidence To ground the ethical analysis in real-world practice, we conducted a structured but non-systematic identification of empirical studies examining the adoption, implementation, or routine use of LLMs in health-care settings. This approach was designed to capture ethically relevant adoption dynamics rather than to provide comprehensive coverage of all published studies. Relevant literature was identified through searches of PubMed/MEDLINE, Scopus, Web of Science, and Google Scholar, supplemented by backward and forward citation tracking. Search terms combined references to LLM technologies (e.g., “large language model*”, “ChatGPT”, “generative AI”, “GPT”) with terms related to real-world use and socio-technical conditions (e.g., “adoption”, “implementation”, “workflow”, “trust”, “communication”, “patient”, “clinician”). We prioritized peer-reviewed empirical studies published between January 2019 and December 2025 that examined LLM use in real or quasi-real health-care settings, with explicit relevance to ethical domains such as trust, responsibility, reliance, governance, or equity. Included studies comprised quantitative, qualitative, and mixed-methods research reporting user experiences, perceptions, or observed patterns of use involving clinicians, patients, and/or caregivers. Purely technical benchmark evaluations, proof-of-concept studies without an adoption component, and non-empirical opinion pieces were excluded. Evidence synthesis and normative interpretation Evidence selection and synthesis were conducted by two authors independently, with discrepancies resolved through discussion and adjudication by a third author when necessary. Rather than extracting standardized outcome measures, we focused on identifying recurring empirical patterns that bear ethical significance—such as shifts in trust formation, ambiguity in responsibility attribution, and differential capacities to benefit from LLM-mediated information. These empirically observed patterns were mapped to normative ethical questions through iterative thematic interpretation, informed by ethical theory and principles relevant to medical practice. The analysis does not treat empirical findings as definitive claims about prevalence or effectiveness; instead, they serve to highlight ethical vulnerabilities and governance gaps that become salient during routine adoption. Supplementary Table S1 summarizes the empirical studies that were analytically central to the ethical themes discussed in this Article, serving to enhance transparency rather than to provide an exhaustive catalogue of the literature. Analytical framework The ethical analysis is structured around three interrelated domains—trust, responsibility, and equity—which consistently emerge as ethically salient during adoption-phase LLM use. Trust is examined as a relational and institutional phenomenon encompassing expectations of transparency, accountability, and protection from harm. Responsibility is analyzed in terms of how LLM integration reshapes professional accountability, decision-making authority, and the allocation of risk across clinicians, institutions, and patients. Equity is considered with respect to differential capacities to engage with and benefit from LLM-mediated information, including variations in digital literacy, educational background, and access to institutional support. These domains are treated as interconnected features of socio-technical systems rather than as isolated ethical principles. Analysis focuses on how their interaction produces ethical tensions that may remain latent during experimental evaluation but become pronounced during routine use. Scope, boundaries, and limitations The scope of this study is limited to LLM applications supporting health-care communication, information interpretation, and non-autonomous decision support, such as explaining medical reports or facilitating patient–clinician dialogue. Fully autonomous clinical decision-making and treatment recommendation systems fall outside the scope of this analysis. This study does not involve human participants or primary data collection and therefore does not require ethics committee approval. Several limitations should be noted. First, reliance on published empirical studies may reflect publication bias and uneven geographic representation. Second, as a normative analysis, the findings do not estimate the prevalence or magnitude of ethical risks. Instead, the contribution lies in clarifying ethical structures and governance needs that warrant further empirical and policy-oriented investigation. To support conceptual clarity, Fig. 1 functions as an analytical heuristic, illustrating how ethical attention shifts from model-level concerns during experimentation to system-level governance issues as LLMs become embedded in routine health care. Ethical implications emerging from routine LLM adoption in health care Trust as an institutional and relational phenomenon Across empirical studies examining real-world adoption of LLMs in health-care settings, a recurring ethical pattern concerns the transformation of trust. Rather than being grounded primarily in confidence in algorithmic accuracy, trust during routine use is shaped by expectations regarding transparency, accountability, and institutional protection[ 9 , 10 , 19 – 22 ]. These expectations become salient once LLMs are embedded into everyday clinical communication rather than evaluated under experimental supervision. For clinicians, adoption-oriented studies consistently indicate that trust in LLM-supported communication depends on whether system use is embedded within governed workflows that clarify liability, oversight, and professional responsibility[ 22 – 26 ]. Where institutional policies remain ambiguous, clinicians report reluctance to rely on LLM-assisted outputs, even when technical performance is perceived as adequate. For patients and caregivers, trust is shaped by whether LLM-mediated information is intelligible, contextualized, and clearly distinguished from professional clinical judgement[ 21 , 23 , 24 , 26 , 27 ]. In contexts where these conditions are not met, engagement with LLM-supported communication diminishes despite high perceived technical capability[ 9 , 10 , 21 , 22 , 27 ]. Taken together, these findings indicate that trust in adoption-phase LLM use is not reducible to individual attitudes toward technology. Instead, trust emerges as a relational and institutional property, dependent on governance signals that specify who stands behind LLM-mediated information and how potential harms are addressed. Responsibility and accountability under mediated information use A second ethically salient pattern emerging from adoption evidence concerns responsibility allocation. As LLMs increasingly mediate access to medical information—through summarization, explanation, or communicative support—empirical studies describe growing ambiguity regarding the locus of responsibility[ 3 , 4 , 25 , 28 , 29 ]. Clinicians remain legally and ethically accountable for patient care, yet often rely on AI-generated content whose provenance, limitations, and uncertainty are not always transparent in routine use. At the same time, patients and caregivers may encounter difficulty distinguishing professional clinical guidance from automated inference. Adoption studies suggest that this ambiguity can shift interpretive and evaluative burdens onto non-professional users, who may lack the expertise or institutional protection required to critically assess LLM-mediated outputs[ 12 , 30 – 32 ]. This configuration reveals a responsibility gap that becomes ethically salient only once LLMs are relied upon in everyday communication, rather than confined to supervised experimental contexts. Importantly, this redistribution of responsibility is not primarily a function of technical performance. Instead, it reflects structural features of how LLMs are integrated into health-care workflows and informational practices, highlighting ethical vulnerabilities that persist even in the absence of overt error. Differential vulnerability and capability amplification A third ethical pattern consistently identified across adoption-oriented studies concerns vulnerability and equity. Engagement with LLM-mediated health information is unevenly distributed, shaped by differences in digital literacy, educational background, and regional economic context[ 9 , 33 , 34 ]. These factors influence not only initial uptake but also users’ capacity to interpret, contextualize, and act upon LLM-generated information. Empirical and conceptual work in this area indicates that individuals with stronger literacy and contextual understanding are more likely to benefit from LLM-enabled communication, while those with fewer resources face compounded barriers and higher risk of misunderstanding[ 35 , 36 ]. Conversely, individuals in economically disadvantaged regions or with limited access to institutional support may face compounded barriers, including reduced confidence in interpreting AI-generated content and fewer opportunities for guided use[ 9 ]. Taken together, these disparities support the interpretation that routine LLM use may function as a capability amplifier : it can magnify existing advantages unless deliberately designed and governed to be literacy-sensitive and equity-oriented[ 35 – 38 ]. Asymmetrical vulnerability across user roles Adoption evidence further reveals that ethical vulnerability is asymmetrically distributed across user roles. Patients and caregivers experience epistemic vulnerability when required to interpret LLM-mediated information without equivalent authority, training, or recourse[ 31 , 39 , 40 ]. Clinicians, by contrast, face distinct risks related to over-reliance or erosion of professional judgement. While both forms of vulnerability are ethically significant, they differ in kind and degree, underscoring that ethical risks associated with routine adoption are not evenly shared. Summary of ethical patterns emerging during adoption Collectively, these findings indicate that routine adoption of LLMs is associated with a shift in ethical risk from model-level concerns toward system-level dynamics involving trust formation, responsibility allocation, and differential vulnerability. These patterns emerge not as isolated implementation challenges, but as structural features of how LLMs become embedded within everyday health-care communication. The ethical significance of these patterns is examined further in the Discussion section, where their implications for existing AI ethics frameworks and governance approaches are considered. Discussion Systematic limitations of existing AI ethics frameworks in the context of LLM adoption The ethical patterns identified in this analysis reveal not merely gaps, but systematic limitations in prevailing AI ethics frameworks when applied to routine LLM adoption in health care. Most existing frameworks were developed during an earlier phase of AI integration, when systems were primarily evaluated as discrete technical tools subject to pre-deployment assessment. Consequently, they have emphasized principles such as transparency, explainability, fairness, and bias mitigation[ 18 ]. While these principles remain necessary, the findings of this study indicate that they are insufficiently specified for the ethical challenges that emerge once LLMs become objects of routine reliance within health-care systems. A central limitation lies in the model-centric orientation of existing frameworks. Transparency is commonly framed as the ability to explain how an algorithm produces outputs, yet adoption-phase evidence suggests that users—both clinicians and patients—are less concerned with internal model logic than with institutional endorsement, accountability, and recourse when harm occurs[ 18 , 30 ]. In routine practice, transparency without accountable governance structures may even contribute to misplaced trust, as users interpret technical openness as a signal of institutional reliability. Similarly, fairness is often operationalized through statistical parity or bias audits, whereas the ethical risks identified in this analysis arise from uneven access to digital literacy, educational resources, regional infrastructure, and trustworthy implementation contexts[ 17 , 20 ]. These forms of inequity are not adequately captured by model-level fairness metrics, yet they directly shape who can meaningfully benefit from LLM-mediated health information. A second, and more fundamental, limitation concerns the individualization of ethical responsibility. Many AI ethics frameworks implicitly assume that informed users—clinicians or patients—can appropriately interpret, contextualize, and contest AI-generated information. The adoption-phase evidence synthesized in this study challenges this assumption. In routine care, time pressure, cognitive load, and asymmetries in expertise significantly constrain meaningful oversight. As LLMs mediate clinical communication, ethical responsibility cannot plausibly rest on individual vigilance alone. Instead, responsibility becomes redistributed across clinicians, institutions, and system designers—often without explicit acknowledgment or governance. This under-specification of responsibility constitutes a structural ethical vulnerability rather than a remediable implementation flaw. Finally, prevailing frameworks tend to treat ethical evaluation as a static, pre-deployment exercise, underestimating the dynamic nature of adoption. Ethical risks evolve as patterns of reliance stabilize, user behavior adapts, and organizational norms shift over time. Frameworks that conceptualize ethics as a checklist applied prior to deployment are therefore ill-equipped to detect or respond to emerging harms that arise only through sustained use[ 14 ]. Taken together, these limitations suggest that existing AI ethics frameworks are not merely incomplete, but misaligned with the ethical realities of routine LLM adoption in health care[ 25 ]. Implications for ethical governance and routine clinical practice If ethical risk emerges primarily during adoption rather than experimentation, ethical responses must likewise shift in focus. First, trust in LLM-supported health-care communication must be institutionally anchored rather than individually negotiated. In practice, users rarely evaluate LLMs solely on technical grounds; instead, they rely on signals of organizational endorsement, accountability, and available recourse. Health-care institutions therefore carry a positive ethical obligation to define how LLM outputs are generated, validated, and monitored within clinical and informational workflows, and to clarify responsibility when AI-mediated information influences communication or decision-making. Second, ethical integration of LLMs requires role-sensitive design and governance. Clinicians and patients interact with LLMs under fundamentally different conditions of authority, expertise, and exposure to harm. For clinicians, ethically appropriate systems must support professional accountability, auditability, and calibrated reliance without increasing cognitive burden. For patients and caregivers, ethical use requires plain-language communication, explicit signaling of uncertainty, and safeguards against over-reliance. Treating these roles as ethically interchangeable risks obscuring asymmetries in vulnerability and undermining trust on both sides of the clinical relationship. Third, equity cannot be treated as a secondary or downstream benefit of technological diffusion. As LLMs increasingly mediate access to medical information, disparities in education, digital capacity, and regional resources translate directly into unequal ethical outcomes. Ethical adoption therefore demands proactive institutional measures, including literacy-sensitive interfaces, adaptive explanatory strategies, and targeted support for underserved populations. Without such measures, LLM integration risks amplifying existing inequities under the guise of informational democratization. Finally, ethical oversight must be understood as a continuous governance process rather than an episodic intervention. Adoption is not a single event but an evolving trajectory shaped by changing patterns of reliance and organizational practice. Health-care systems must therefore establish mechanisms for ongoing evaluation, feedback, and revision of LLM deployment, recognizing that ethical risks may emerge, intensify, or transform over time. Together, these considerations point toward an ethics of adoption grounded in institutional responsibility, inclusivity, and sustained oversight rather than isolated technical safeguards. Conclusion Large language models are no longer confined to experimental settings; they are increasingly shaping how medical information is accessed, interpreted, and acted upon in everyday health care. This transition from experimentation to routine adoption constitutes not merely a technological milestone, but a normatively consequential ethical shift. This Article demonstrates that the most pressing ethical challenges associated with LLMs arise less from their technical capabilities than from the ways in which they become embedded within health-care systems, institutional workflows, and social relationships of care. Our analysis shows that routine adoption systematically reconfigures trust, redistributes responsibility, and exposes asymmetrical vulnerability across user groups. These ethical tensions cannot be adequately addressed through model-level improvements alone. In the absence of institutional accountability, role-sensitive design, and equity-oriented governance, LLM integration risks producing ethically problematic outcomes—such as misplaced trust, responsibility gaps, and the amplification of existing disparities—even when technical performance is high. Conversely, when embedded within robust governance structures, LLMs can support clearer communication, enhance understanding, and expand access to health information without undermining professional responsibility or patient protection. The ethical task ahead is therefore not to restrain innovation, but to take adoption seriously as an ongoing ethical process rather than a one-time technical achievement. Health-care institutions, researchers, and policymakers must recognize that ethical responsibility does not end at deployment, but persists as patterns of reliance stabilize and evolve over time. By shifting ethical attention from experimental validation toward system-level governance and longitudinal oversight, health-care systems can better ensure that the expanding use of LLMs advances not only efficiency and scale, but also fairness, accountability, and patient trust. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Availability of data and materials Data sharing is not applicable to this article as no datasets were generated or analyzed during the current study. Competing interests The authors declare that they have no competing interests. Funding This work was supported by the Talent Fund of Guizhou Provincial People’s Hospital (Grant No. 2024-16) and the Science and Technology Fund Project of the Guizhou Provincial Health Commission (Grant No. 2026GZWJKJXM0317). The funding sources had no role in the study design, analysis, interpretation of data, writing of the manuscript, or the decision to submit the manuscript for publication. Authors’ contributions XY, YW, and JH conceptualized and designed the study. XY led the methodological development and formal ethical analysis. XY, YL, YJL, DL, LY, BZ, and CX contributed to data curation and analysis of relevant empirical literature and adoption studies. XY, YL, YJL, DL, LY, BZ, and CX drafted the initial manuscript. YW and JH critically reviewed and revised the manuscript for important intellectual content and provided overall supervision. YW and JH were responsible for project administration and funding acquisition. All authors read and approved the final manuscript. Acknowledgements The authors used the generative AI tool GPT-5 (OpenAI, San Francisco, CA, USA) for language refinement, grammar correction, and consistency editing of the English manuscript. The tool was not used for conceptualization, ethical analysis, interpretation, or reference generation. All scientific content, analysis, and conclusions were developed, verified, and approved by the authors. Authors’ information Not applicable. References Menezes MCS, Hoffmann AF, Tan ALM, Nalbandyan M, Omenn GS, Mazzotti DR, et al. The potential of generative pre-trained transformer 4 (GPT-4) to analyse medical notes in three different languages: a retrospective model-evaluation study. Lancet Digit Health. 2025;7(1):e35–43. 10.1016/S2589-7500(24)00246-2 . Sun Z, Ong H, Kennedy P, Tang L, Chen S, Elias J, et al. Evaluating GPT4 on Impressions Generation in Radiology Reports. Radiology. 2023;307(5):e231259. 10.1148/radiol.231259 . Thirunavukarasu AJ, Ting DSJ, Elangovan K, Gutierrez L, Tan TF, Ting DSW. Large language models in medicine. Nat Med. 2023;29(8):1930–40. 10.1038/s41591-023-02448-8 . Singhal K, Azizi S, Tu T, Mahdavi SS, Wei J, Chung HW, et al. Large language models encode clinical knowledge. Nature. 2023;620(7972):172–80. 10.1038/s41586-023-06291-2 . Yang X, Xiao Y, Liu D, Deng H, Huang J, Zhou Y, et al. Cross language transformation of free text into structured lobectomy surgical records from a multi center study. Sci Rep. 2025;15(1):15417. 10.1038/s41598-025-97500-7 . Yang X, Zhang Y, Jiang J, Chen Z, Bai R, Yuan Z, et al. Harnessing GPT-4 for automated error detection in pathology reports: Implications for oncology diagnostics. Digit Health. 2025;11:20552076251346703. 10.1177/20552076251346703 . Yang X, Xiao Y, Liu D, Zhang Y, Deng H, Huang J, et al. Enhancing doctor-patient communication using large language models for pathology report interpretation. BMC Med Inf Decis Mak. 2025;25(1):36. 10.1186/s12911-024-02838-z . Yang X, Xiao Y, Liu D, Shi H, Deng H, Huang J, et al. Enhancing Physician-Patient Communication in Oncology Using GPT-4 Through Simplified Radiology Reports: Multicenter Quantitative Study. J Med Internet Res. 2025;27:e63786. 10.2196/63786 . Yang X, Xiao Y, Liu D, Deng H, Huang J, Zhou Y, et al. Factors Influencing Adoption of Large Language Models in Health Care: Multicenter Cross-Sectional Mixed Methods Observational Study. J Med Internet Res. 2025;27:e84918. 10.2196/84918 . Chen J, Liu Y, Liu P, Zhao Y, Zuo Y, Duan H. Adoption of Large Language Model AI Tools in Everyday Tasks: Multisite Cross-Sectional Qualitative Study of Chinese Hospital Administrators. J Med Internet Res. 2025;27:e70789. 10.2196/70789 . Dennstädt F, Schmerder M, Riggenbach E, Mose L, Bryjova K, Bachmann N, et al. Comparative Evaluation of a Medical Large Language Model in Answering Real-World Radiation Oncology Questions: Multicenter Observational Study. J Med Internet Res. 2025;27:e69752. 10.2196/69752 . Shin HS, Williams H, Braykov N, Jahan A, Meller J, Orenstein EW. The Influence of Artificial Intelligence Scribes on Clinician Experience and Efficiency among Pediatric Subspecialists: A Rapid, Randomized Quality Improvement Trial. Appl Clin Inf. 2025;16(4):1041–52. 10.1055/a-2657-8087 . Hanna MG, Pantanowitz L, Jackson B, Palmer O, Visweswaran S, Pantanowitz J, et al. Ethical and Bias Considerations in Artificial Intelligence/Machine Learning. Mod Pathol. 2025;38(3):100686. https://doi.org/10.1016/j.modpat.2024.100686 . Torkamaan H, Steinert S, Pera MS, Kudina O, Freire SK, Verma H, et al. Challenges and future directions for integration of large language models into socio-technical systems. Behav Inform Technol. 2024;1–20. 10.1080/0144929X.2024.2431068 . Dhagarra D, Goswami M, Kumar G. Impact of Trust and Privacy Concerns on Technology Acceptance in Healthcare: An Indian Perspective. Int J Med Inf. 2020;141:104164. 10.1016/j.ijmedinf.2020.104164 . Kaur D, Uslu S, Rittichier KJ, Durresi A. Trustworthy Artificial Intelligence: A Review. ACM Comput Surv. 2022;55(2).):Article 39. 10.1145/3491209 . Singhal A, Neveditsin N, Tanveer H, Mago V, Toward Fairness. Accountability, Transparency, and Ethics in AI for Social Media and Health Care: Scoping Review. JMIR Med Inf. 2024;12:e50048. 10.2196/50048 . Vollmer S, Mateen BA, Bohner G, Király FJ, Ghani R, Jonsson P, et al. Machine learning and artificial intelligence research for patient benefit: 20 critical questions on transparency, replicability, ethics, and effectiveness. BMJ. 2020;368:l6927. 10.1136/bmj.l6927 . Massenon R, Gambo I, Khan JA, Agbonkhese C, Alwadain A. My AI is Lying to Me: User-reported LLM hallucinations in AI mobile apps reviews. Sci Rep. 2025;15(1):30397. 10.1038/s41598-025-15416-8 . Visser R, Peters TM, Scharlau I, Hammer B. Trust, distrust, and appropriate reliance in (X)AI: A conceptual clarification of user trust and survey of its empirical evaluation. Cogn Syst Res. 2025;91:101357. https://doi.org/10.1016/j.cogsys.2025.101357 . Kim SH, Wihl J, Schramm S, Berberich C, Rosenkranz E, Schmitzer L, et al. Human-AI collaboration in large language model-assisted brain MRI differential diagnosis: a usability study. Eur Radiol. 2025;35(9):5252–63. 10.1007/s00330-025-11484-6 . Ebnali Harari R, Altaweel A, Ahram T, Keehner M, Shokoohi H. A randomized controlled trial on evaluating clinician-supervised generative AI for decision support. Int J Med Inf. 2025;195:105701. 10.1016/j.ijmedinf.2024.105701 . Allen LN, Lin J, Segal BM, Ndlovu K, Bilardi D, Pettigrew LM. Artificial intelligence in primary care: frameworks, challenges, and guardrails. Lancet Prim Care. 10.1016/j.lanprc.2025.100079 Heston TF, Lu M, Gillette J. Ethical Hazards of Large Language Models in Primary Care: A Clinician-Focused Update. In: Heston TF, editor. Medical Ethics - Navigating Complex Decisions in Contemporary Healthcare. London: IntechOpen; 2025. Ryan K, Kasun M, Roberts LW, Kim JP. Information, collaboration, regulation: Physician and AI researcher views on ethical considerations in clinical AI integration. Big Data Soc. 2025;12(2):20539517251343853. 10.1177/20539517251343853 . Yıldırım C, Aykut A, Günsoy E, Öncül MV. Evaluating GPT-4o for emergency disposition of complex respiratory cases with pulmonology consultation: a diagnostic accuracy study. Scand J Trauma Resusc Emerg Med. 2025;33(1):159. 10.1186/s13049-025-01475-3 . Lee S, Jung S, Park JH, Cho H, Moon S, Ahn S. Performance of ChatGPT, Gemini and DeepSeek for non-critical triage support using real-world conversations in emergency department. BMC Emerg Med. 2025;25(1):176. 10.1186/s12873-025-01337-2 . Azevedo CB, Martinho AS, Braga I, Nogueira-Silva C, Barroso C, Correia-Pinto J. ChatGPT-4o in Enhancing Informed Consent in Pediatric Surgical Practice. J Pediatr Surg. 2025;60(9):162413. 10.1016/j.jpedsurg.2025.162413 . Umman V, Tosun B, Uygur A, Emre S. Evaluation of the Effectiveness, Safety, and Patient Satisfaction of Artificial Intelligence-Based Patient Education and Counseling for Both Recipients and Donors in the Preoperative and Postoperative Phases of Organ Transplantation. Transplant Proc. 2025;57(9):1832-9. 10.1016/j.transproceed.2025.07.001 Nouis SC, Uren V, Jariwala S. Evaluating accountability, transparency, and bias in AI-assisted healthcare decision- making: a qualitative study of healthcare professionals' perspectives in the UK. BMC Med Ethics. 2025;26(1):89. 10.1186/s12910-025-01243-z . Fink A, Nattenmüller J, Rau S, Rau A, Tran H, Bamberg F, et al. Retrieval-augmented generation improves precision and trust of a GPT-4 model for emergency radiology diagnosis and classification: a proof-of-concept study. Eur Radiol. 2025;35(8):5091–8. 10.1007/s00330-025-11445-z . Fukui Y, Kawata Y, Kobashi K, Nagatani Y, Iguchi H. Evaluation of a retrieval-augmented generation system using a Japanese Institutional Nuclear Medicine Manual and large language model-automated scoring. Radiol Phys Technol. 2025;18(3):861–76. 10.1007/s12194-025-00941-y . Dang Q, Li G. Unveiling trust in AI: the interplay of antecedents, consequences, and cultural dynamics. AI Soc. 2025. 10.1007/s00146-025-02477-6 . Urbina JT, Vu PD, Nguyen MV. Disability Ethics and Education in the Age of Artificial Intelligence: Identifying Ability Bias in ChatGPT and Gemini. Arch Phys Med Rehabil. 2025;106(1):14–9. 10.1016/j.apmr.2024.08.014 . Ji Y, Ma W, Sivarajkumar S, Zhang H, Sadhu EM, Li Z, et al. Mitigating the risk of health inequity exacerbated by large language models. npj Digit Med. 2025;8(1):246. 10.1038/s41746-025-01576-4 . Naghdi M, Cao P, Essers R, Heijligers M, Paulussen ADC, van der Lugt A, et al. Artificial intelligence-simplified information to advance reproductive genetic literacy and health equity. Hum Reprod. 2025;40(9):1681–8. 10.1093/humrep/deaf135 . Ahmed H. Large language models for clinical trials in the Global South: opportunities and ethical challenges. AI Ethics. 2025;6(1):76. 10.1007/s43681-025-00943-x . Rodler S, Cei F, Ganjavi C, Checcucci E, De Backer P, Rivero Belenchon I, et al. GPT-4 generates accurate and readable patient education materials aligned with current oncological guidelines: A randomized assessment. PLoS ONE. 2025;20(6):e0324175. 10.1371/journal.pone.0324175 . Campellone TR, Flom M, Montgomery RM, Bullard L, Pirner MC, Pavez A, et al. Safety and User Experience of a Generative Artificial Intelligence Digital Mental Health Intervention: Exploratory Randomized Controlled Trial. J Med Internet Res. 2025;27:e67365. 10.2196/67365 . Liu Y, Calle P, Vadakekut M, Rubin D, Nagykaldi Z, Doescher M, et al. AI-Enabled Personalized Smoking Cessation Intervention With the Aipaca Chatbot: Mixed Methods Feasibility Study. JMIR Form Res. 2025;9:e73319. 10.2196/73319 . Additional Declarations No competing interests reported. Supplementary Files TableS1.xlsx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version 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-8553262","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":577923566,"identity":"cd653a56-4b29-4596-ae17-737fa7ec6b62","order_by":0,"name":"Xiongwen Yang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABBklEQVRIiWNgGAWjYNACAyj9gYGhnk3+/APitTDOYGBI4JfgIcEyZqDiBMkZBLQYHD97+DVPgU2efETyscc2v+ryDG73HmD4UbENt5YzeWmWMwzSig1vpKUb5/axFRvcOZfA2HPmNm4tB3LMDD4YHE7cODvHTDq3h4dxw4EEA2bGNjxazr8xM0gAa8n/Jm3ZI0GElhs5xg9AtsyXzmGTZvhhkDhzRg5+LZI33pgxAv2SuEH+mZlkb0OCMT/PsYSD+PzCdz7H+DPPH5vE+T2Hn0n8+FMnx8befPDBjwrcWhQOMLBJQMIBSDC2QUQP4FQPBPINDMwfoAwg+INP7SgYBaNgFIxUAABT01+Rru8vzwAAAABJRU5ErkJggg==","orcid":"","institution":"Guizhou Provincial People's Hospital","correspondingAuthor":true,"prefix":"","firstName":"Xiongwen","middleName":"","lastName":"Yang","suffix":""},{"id":577923567,"identity":"73cb497a-a0fe-4820-938f-97dcbdab462a","order_by":1,"name":"Yuanwei Liang","email":"","orcid":"","institution":"Third Affiliated Hospital of Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Yuanwei","middleName":"","lastName":"Liang","suffix":""},{"id":577923568,"identity":"0e3f2cd2-d444-42ce-b9c4-b5dcb8ae191c","order_by":2,"name":"Yijiang Liu","email":"","orcid":"","institution":"Hainan Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yijiang","middleName":"","lastName":"Liu","suffix":""},{"id":577923569,"identity":"744d7001-7d37-438b-a058-3e6b16251401","order_by":3,"name":"Di Liu","email":"","orcid":"","institution":"Guizhou Provincial People's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Di","middleName":"","lastName":"Liu","suffix":""},{"id":577923570,"identity":"8e17fd4a-a0a1-46c7-918f-49601681e524","order_by":4,"name":"Lin Yang","email":"","orcid":"","institution":"Guizhou Provincial People's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Lin","middleName":"","lastName":"Yang","suffix":""},{"id":577923571,"identity":"d4374ce5-7d14-414d-9d3a-6015c6bb684f","order_by":5,"name":"Bo Zhang","email":"","orcid":"","institution":"Guizhou Provincial People's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Bo","middleName":"","lastName":"Zhang","suffix":""},{"id":577923572,"identity":"e392e93d-f81e-4645-8a51-cda62ed8d947","order_by":6,"name":"Chuan Xu","email":"","orcid":"","institution":"Guizhou Provincial People's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Chuan","middleName":"","lastName":"Xu","suffix":""},{"id":577923573,"identity":"cdfde2c8-e2d1-4085-a581-022ec3c4f68c","order_by":7,"name":"YongHui Wu","email":"","orcid":"","institution":"Third Affiliated Hospital of Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"YongHui","middleName":"","lastName":"Wu","suffix":""},{"id":577923574,"identity":"e5f0aa75-f52f-4a27-bab9-e24c45128035","order_by":8,"name":"Jinyuan He","email":"","orcid":"","institution":"Third Affiliated Hospital of Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Jinyuan","middleName":"","lastName":"He","suffix":""}],"badges":[],"createdAt":"2026-01-08 15:39:09","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8553262/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8553262/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":100929564,"identity":"8a56d7dc-6202-4f1f-b992-437a8326a34d","added_by":"auto","created_at":"2026-01-23 00:36:52","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":149318,"visible":true,"origin":"","legend":"","description":"","filename":"ethicalchallengesBMC.docx","url":"https://assets-eu.researchsquare.com/files/rs-8553262/v1/39e73950e27544bb548585aa.docx"},{"id":100929559,"identity":"fde103f1-2934-4843-a52f-1d5a20687972","added_by":"auto","created_at":"2026-01-23 00:36:52","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10661,"visible":true,"origin":"","legend":"","description":"","filename":"ac498e87ccc14cccab5fa43f2466ce75.json","url":"https://assets-eu.researchsquare.com/files/rs-8553262/v1/e42611489fd4b6fa4c8d953b.json"},{"id":100929562,"identity":"96eb4fa4-f503-4a32-b8db-5ea532d95299","added_by":"auto","created_at":"2026-01-23 00:36:52","extension":"xlsx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":22208,"visible":true,"origin":"","legend":"","description":"","filename":"TableS1.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8553262/v1/3586c4f5b9f785e4d3977955.xlsx"},{"id":100929565,"identity":"d388617e-c563-43a1-b91b-4a8fa6d64aa8","added_by":"auto","created_at":"2026-01-23 00:36:52","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":108428,"visible":true,"origin":"","legend":"","description":"","filename":"ac498e87ccc14cccab5fa43f2466ce751enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8553262/v1/30901b182daee6f20e368f97.xml"},{"id":100929567,"identity":"71af7735-764c-41a0-ae5d-2f2f9cbcfa51","added_by":"auto","created_at":"2026-01-23 00:36:52","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":35598,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8553262/v1/4556eac7e5472ead9e54610c.png"},{"id":100951261,"identity":"15eef66a-268c-47fc-ba5a-31b611c9ad19","added_by":"auto","created_at":"2026-01-23 07:10:20","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":17741,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8553262/v1/99d4999aef61b8669a23e1fc.png"},{"id":101296640,"identity":"f9a9823e-a754-4ceb-8fa1-e79dbe2c39b7","added_by":"auto","created_at":"2026-01-28 09:17:42","extension":"xml","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":108061,"visible":true,"origin":"","legend":"","description":"","filename":"ac498e87ccc14cccab5fa43f2466ce751structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8553262/v1/9d7161d808429b248792f9dd.xml"},{"id":100929568,"identity":"5cc329a7-b5ab-43c8-8616-c483ac50b440","added_by":"auto","created_at":"2026-01-23 00:36:52","extension":"html","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":118929,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8553262/v1/89b9b8de90dee9150538c6bc.html"},{"id":100929558,"identity":"8f5c633b-64af-4f41-884d-029d2308ebac","added_by":"auto","created_at":"2026-01-23 00:36:52","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":47921,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEthical shift from experimentation to adoption of large language models in health care.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis schematic illustrates how the ethical focus of large language models (LLMs) in health care evolves as these systems move from experimental evaluation to routine adoption. During experimentation, ethical considerations are primarily situated at the model level, including accuracy, validation, and bias mitigation. As LLMs are increasingly integrated into clinical workflows and health-care communication involving clinicians, patients, and caregivers, ethical attention shifts toward system-level issues. At the adoption stage, ethical challenges are predominantly related to trust, responsibility, and equity, highlighting the need for governance approaches that extend beyond isolated technical safeguards. This figure is conceptual and serves as an analytical heuristic rather than an empirical model.\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8553262/v1/0a24a01547d4a6f2c359f261.jpg"},{"id":102376284,"identity":"f7608097-8b30-4203-95e4-ff8d728315de","added_by":"auto","created_at":"2026-02-11 05:26:51","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":888997,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8553262/v1/a5f0ad16-af66-4c34-b266-bd80ebfb60c1.pdf"},{"id":100951781,"identity":"f617bb59-33c3-4bb5-b1a5-e707eb661497","added_by":"auto","created_at":"2026-01-23 07:11:14","extension":"xlsx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":22208,"visible":true,"origin":"","legend":"","description":"","filename":"TableS1.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8553262/v1/3b3d43f0f86d52f8060e120f.xlsx"}],"financialInterests":"No competing interests reported.","formattedTitle":"From experimentation to adoption: a normative ethical analysis of large language models in health care","fulltext":[{"header":"Introduction","content":"\u003cp\u003eLarge language models (LLMs) have rapidly moved from experimental tools to increasingly visible actors within health-care information ecosystems. Early research has primarily examined whether these systems can perform clinically relevant tasks\u0026mdash;such as summarizing medical reports, explaining technical terminology, or supporting decision-making\u0026mdash;with acceptable accuracy and efficiency[\u003cspan additionalcitationids=\"CR2 CR3 CR4 CR5\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Such work has been essential in establishing proof of concept and translational potential. However, as LLMs are increasingly adopted by clinicians, patients, and caregivers in real-world settings[\u003cspan additionalcitationids=\"CR8 CR9 CR10 CR11\" citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], a distinct ethical question arises that cannot be answered by performance evaluation alone: what changes ethically when LLMs become objects of routine reliance in everyday care, rather than instruments tested under experimental supervision?\u003c/p\u003e \u003cp\u003eWe argue that the shift from experimentation to adoption marks an adoption-phase ethics problem: ethical risk increasingly emerges not from isolated model outputs, but from the ways LLMs are embedded into clinical relationships, institutional workflows, and expectations of care. During experimental use, ethical concerns are commonly framed in technical terms\u0026mdash;including model performance, validation, and bias mitigation[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. In contrast, routine adoption redistributes epistemic authority over medical information, reshapes how trust is negotiated between clinicians and patients, and complicates established lines of professional responsibility[\u003cspan additionalcitationids=\"CR10 CR11\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. These changes are not merely conceptual. Clinicians remain legally and ethically accountable for clinical decisions, yet patients and caregivers may increasingly act on LLM-mediated explanations when interpreting diagnoses, prognoses, and treatment options[\u003cspan additionalcitationids=\"CR10 CR11\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. This configuration creates an ethics of reliance problem: reliance can grow faster than the capacity of users and institutions to interpret uncertainty, contest outputs, or secure recourse when harm occurs.\u003c/p\u003e \u003cp\u003eEmerging evidence from adoption-oriented studies reinforces that sustained uptake of LLMs in health care is driven less by algorithmic sophistication than by socio-technical and ethical conditions[\u003cspan additionalcitationids=\"CR10 CR11\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Trust, perceived usefulness, digital literacy, privacy assurance, and institutional support consistently shape whether and how LLMs are used[\u003cspan additionalcitationids=\"CR10 CR11\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Importantly, trust here is not reducible to confidence in accuracy; it reflects expectations about who stands behind LLM-mediated information, how limitations are signaled, and what protections exist when expectations are violated[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Where such governance signals are weak, adoption may stall\u0026mdash;or proceed in informal, uneven, and potentially unsafe ways\u0026mdash;particularly among individuals with lower digital literacy or limited access to institutional guidance, raising concerns about inequity and exclusion.\u003c/p\u003e \u003cp\u003eThese dynamics expose limits in prevailing approaches to AI ethics in medicine, which often assume that improving model performance, explainability, or transparency is sufficient to manage ethical risk[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. As LLMs move into routine clinical and informational use, ethical hazards increasingly arise from institutional embedding rather than from technical properties alone. Transparency without accountable workflows can still produce misplaced trust; explainability without role-appropriate safeguards can still shift burdens onto patients and front-line clinicians. Ethical attention must therefore expand beyond pre-deployment evaluation toward system-level governance questions about trust, responsibility allocation, and equitable capability to benefit.\u003c/p\u003e \u003cp\u003eIn this Article, adoption is understood not merely as technical deployment, but as the sustained integration of LLMs into everyday clinical and informational practices through which patterns of reliance, trust, and responsibility become socially and institutionally stabilized[\u003cspan additionalcitationids=\"CR10 CR11\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Using a normative ethical analysis informed by empirical adoption evidence, we examine three interrelated domains\u0026mdash;trust, responsibility, and equity\u0026mdash;to clarify how routine LLM use can (i) reconfigure institutional and interpersonal trust, (ii) produce responsibility and accountability gaps, and (iii) amplify existing disparities in the capacity to interpret and benefit from AI-mediated information. We focus on health-care communication and non-autonomous decision support rather than fully autonomous diagnosis or treatment recommendation. We conclude by outlining governance implications for ethically robust integration of LLMs into health-care communication.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStudy design and ethical methodology\u003c/h2\u003e \u003cp\u003eThis study employs a normative ethical analysis informed by empirical evidence, focusing on ethical challenges that arise during the adoption phase of LLMs in health care. The aim is not to evaluate technical performance or quantify adoption outcomes, but to identify ethically salient patterns that emerge when LLMs move from experimental evaluation to routine use in clinical communication and informational contexts.\u003c/p\u003e \u003cp\u003eConsistent with established approaches in medical ethics, empirical findings are used as contextual inputs to normative reasoning rather than as sources of causal inference. Empirical evidence is mobilized to illuminate how trust relationships, responsibility allocation, and equity conditions are reconfigured through sustained reliance on LLM-mediated information.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eIdentification and selection of empirical evidence\u003c/h3\u003e\n\u003cp\u003eTo ground the ethical analysis in real-world practice, we conducted a structured but non-systematic identification of empirical studies examining the adoption, implementation, or routine use of LLMs in health-care settings. This approach was designed to capture ethically relevant adoption dynamics rather than to provide comprehensive coverage of all published studies.\u003c/p\u003e \u003cp\u003eRelevant literature was identified through searches of PubMed/MEDLINE, Scopus, Web of Science, and Google Scholar, supplemented by backward and forward citation tracking. Search terms combined references to LLM technologies (e.g., \u0026ldquo;large language model*\u0026rdquo;, \u0026ldquo;ChatGPT\u0026rdquo;, \u0026ldquo;generative AI\u0026rdquo;, \u0026ldquo;GPT\u0026rdquo;) with terms related to real-world use and socio-technical conditions (e.g., \u0026ldquo;adoption\u0026rdquo;, \u0026ldquo;implementation\u0026rdquo;, \u0026ldquo;workflow\u0026rdquo;, \u0026ldquo;trust\u0026rdquo;, \u0026ldquo;communication\u0026rdquo;, \u0026ldquo;patient\u0026rdquo;, \u0026ldquo;clinician\u0026rdquo;).\u003c/p\u003e \u003cp\u003eWe prioritized peer-reviewed empirical studies published between January 2019 and December 2025 that examined LLM use in real or quasi-real health-care settings, with explicit relevance to ethical domains such as trust, responsibility, reliance, governance, or equity. Included studies comprised quantitative, qualitative, and mixed-methods research reporting user experiences, perceptions, or observed patterns of use involving clinicians, patients, and/or caregivers. Purely technical benchmark evaluations, proof-of-concept studies without an adoption component, and non-empirical opinion pieces were excluded.\u003c/p\u003e\n\u003ch3\u003eEvidence synthesis and normative interpretation\u003c/h3\u003e\n\u003cp\u003eEvidence selection and synthesis were conducted by two authors independently, with discrepancies resolved through discussion and adjudication by a third author when necessary. Rather than extracting standardized outcome measures, we focused on identifying recurring empirical patterns that bear ethical significance\u0026mdash;such as shifts in trust formation, ambiguity in responsibility attribution, and differential capacities to benefit from LLM-mediated information.\u003c/p\u003e \u003cp\u003eThese empirically observed patterns were mapped to normative ethical questions through iterative thematic interpretation, informed by ethical theory and principles relevant to medical practice. The analysis does not treat empirical findings as definitive claims about prevalence or effectiveness; instead, they serve to highlight ethical vulnerabilities and governance gaps that become salient during routine adoption. \u003cb\u003eSupplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e\u003c/b\u003e summarizes the empirical studies that were analytically central to the ethical themes discussed in this Article, serving to enhance transparency rather than to provide an exhaustive catalogue of the literature.\u003c/p\u003e\n\u003ch3\u003eAnalytical framework\u003c/h3\u003e\n\u003cp\u003eThe ethical analysis is structured around three interrelated domains\u0026mdash;trust, responsibility, and equity\u0026mdash;which consistently emerge as ethically salient during adoption-phase LLM use. Trust is examined as a relational and institutional phenomenon encompassing expectations of transparency, accountability, and protection from harm. Responsibility is analyzed in terms of how LLM integration reshapes professional accountability, decision-making authority, and the allocation of risk across clinicians, institutions, and patients. Equity is considered with respect to differential capacities to engage with and benefit from LLM-mediated information, including variations in digital literacy, educational background, and access to institutional support.\u003c/p\u003e \u003cp\u003eThese domains are treated as interconnected features of socio-technical systems rather than as isolated ethical principles. Analysis focuses on how their interaction produces ethical tensions that may remain latent during experimental evaluation but become pronounced during routine use.\u003c/p\u003e\n\u003ch3\u003eScope, boundaries, and limitations\u003c/h3\u003e\n\u003cp\u003eThe scope of this study is limited to LLM applications supporting health-care communication, information interpretation, and non-autonomous decision support, such as explaining medical reports or facilitating patient\u0026ndash;clinician dialogue. Fully autonomous clinical decision-making and treatment recommendation systems fall outside the scope of this analysis.\u003c/p\u003e \u003cp\u003eThis study does not involve human participants or primary data collection and therefore does not require ethics committee approval. Several limitations should be noted. First, reliance on published empirical studies may reflect publication bias and uneven geographic representation. Second, as a normative analysis, the findings do not estimate the prevalence or magnitude of ethical risks. Instead, the contribution lies in clarifying ethical structures and governance needs that warrant further empirical and policy-oriented investigation.\u003c/p\u003e \u003cp\u003eTo support conceptual clarity, Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e functions as an analytical heuristic, illustrating how ethical attention shifts from model-level concerns during experimentation to system-level governance issues as LLMs become embedded in routine health care.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eEthical implications emerging from routine LLM adoption in health care\u003c/h2\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003eTrust as an institutional and relational phenomenon\u003c/h2\u003e \u003cp\u003eAcross empirical studies examining real-world adoption of LLMs in health-care settings, a recurring ethical pattern concerns the transformation of trust. Rather than being grounded primarily in confidence in algorithmic accuracy, trust during routine use is shaped by expectations regarding transparency, accountability, and institutional protection[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan additionalcitationids=\"CR20 CR21\" citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. These expectations become salient once LLMs are embedded into everyday clinical communication rather than evaluated under experimental supervision.\u003c/p\u003e \u003cp\u003eFor clinicians, adoption-oriented studies consistently indicate that trust in LLM-supported communication depends on whether system use is embedded within governed workflows that clarify liability, oversight, and professional responsibility[\u003cspan additionalcitationids=\"CR23 CR24 CR25\" citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. Where institutional policies remain ambiguous, clinicians report reluctance to rely on LLM-assisted outputs, even when technical performance is perceived as adequate. For patients and caregivers, trust is shaped by whether LLM-mediated information is intelligible, contextualized, and clearly distinguished from professional clinical judgement[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. In contexts where these conditions are not met, engagement with LLM-supported communication diminishes despite high perceived technical capability[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTaken together, these findings indicate that trust in adoption-phase LLM use is not reducible to individual attitudes toward technology. Instead, trust emerges as a relational and institutional property, dependent on governance signals that specify who stands behind LLM-mediated information and how potential harms are addressed.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e\n\u003ch3\u003eResponsibility and accountability under mediated information use\u003c/h3\u003e\n\u003cp\u003eA second ethically salient pattern emerging from adoption evidence concerns responsibility allocation. As LLMs increasingly mediate access to medical information\u0026mdash;through summarization, explanation, or communicative support\u0026mdash;empirical studies describe growing ambiguity regarding the locus of responsibility[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. Clinicians remain legally and ethically accountable for patient care, yet often rely on AI-generated content whose provenance, limitations, and uncertainty are not always transparent in routine use.\u003c/p\u003e \u003cp\u003eAt the same time, patients and caregivers may encounter difficulty distinguishing professional clinical guidance from automated inference. Adoption studies suggest that this ambiguity can shift interpretive and evaluative burdens onto non-professional users, who may lack the expertise or institutional protection required to critically assess LLM-mediated outputs[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan additionalcitationids=\"CR31\" citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. This configuration reveals a responsibility gap that becomes ethically salient only once LLMs are relied upon in everyday communication, rather than confined to supervised experimental contexts.\u003c/p\u003e \u003cp\u003eImportantly, this redistribution of responsibility is not primarily a function of technical performance. Instead, it reflects structural features of how LLMs are integrated into health-care workflows and informational practices, highlighting ethical vulnerabilities that persist even in the absence of overt error.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eDifferential vulnerability and capability amplification\u003c/h2\u003e \u003cp\u003eA third ethical pattern consistently identified across adoption-oriented studies concerns vulnerability and equity. Engagement with LLM-mediated health information is unevenly distributed, shaped by differences in digital literacy, educational background, and regional economic context[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. These factors influence not only initial uptake but also users\u0026rsquo; capacity to interpret, contextualize, and act upon LLM-generated information.\u003c/p\u003e \u003cp\u003eEmpirical and conceptual work in this area indicates that individuals with stronger literacy and contextual understanding are more likely to benefit from LLM-enabled communication, while those with fewer resources face compounded barriers and higher risk of misunderstanding[\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. Conversely, individuals in economically disadvantaged regions or with limited access to institutional support may face compounded barriers, including reduced confidence in interpreting AI-generated content and fewer opportunities for guided use[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Taken together, these disparities support the interpretation that routine LLM use may function as a \u003cem\u003ecapability amplifier\u003c/em\u003e: it can magnify existing advantages unless deliberately designed and governed to be literacy-sensitive and equity-oriented[\u003cspan additionalcitationids=\"CR36 CR37\" citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eAsymmetrical vulnerability across user roles\u003c/h2\u003e \u003cp\u003eAdoption evidence further reveals that ethical vulnerability is asymmetrically distributed across user roles. Patients and caregivers experience epistemic vulnerability when required to interpret LLM-mediated information without equivalent authority, training, or recourse[\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e, \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e]. Clinicians, by contrast, face distinct risks related to over-reliance or erosion of professional judgement. While both forms of vulnerability are ethically significant, they differ in kind and degree, underscoring that ethical risks associated with routine adoption are not evenly shared.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eSummary of ethical patterns emerging during adoption\u003c/h2\u003e \u003cp\u003eCollectively, these findings indicate that routine adoption of LLMs is associated with a shift in ethical risk from model-level concerns toward system-level dynamics involving trust formation, responsibility allocation, and differential vulnerability. These patterns emerge not as isolated implementation challenges, but as structural features of how LLMs become embedded within everyday health-care communication. The ethical significance of these patterns is examined further in the Discussion section, where their implications for existing AI ethics frameworks and governance approaches are considered.\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eSystematic limitations of existing AI ethics frameworks in the context of LLM adoption\u003c/h2\u003e \u003cp\u003eThe ethical patterns identified in this analysis reveal not merely gaps, but \u003cb\u003esystematic limitations\u003c/b\u003e in prevailing AI ethics frameworks when applied to routine LLM adoption in health care. Most existing frameworks were developed during an earlier phase of AI integration, when systems were primarily evaluated as discrete technical tools subject to pre-deployment assessment. Consequently, they have emphasized principles such as transparency, explainability, fairness, and bias mitigation[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. While these principles remain necessary, the findings of this study indicate that they are insufficiently specified for the ethical challenges that emerge once LLMs become objects of routine reliance within health-care systems.\u003c/p\u003e \u003cp\u003eA central limitation lies in the model-centric orientation of existing frameworks. Transparency is commonly framed as the ability to explain how an algorithm produces outputs, yet adoption-phase evidence suggests that users\u0026mdash;both clinicians and patients\u0026mdash;are less concerned with internal model logic than with institutional endorsement, accountability, and recourse when harm occurs[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. In routine practice, transparency without accountable governance structures may even contribute to misplaced trust, as users interpret technical openness as a signal of institutional reliability. Similarly, fairness is often operationalized through statistical parity or bias audits, whereas the ethical risks identified in this analysis arise from uneven access to digital literacy, educational resources, regional infrastructure, and trustworthy implementation contexts[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. These forms of inequity are not adequately captured by model-level fairness metrics, yet they directly shape who can meaningfully benefit from LLM-mediated health information.\u003c/p\u003e \u003cp\u003eA second, and more fundamental, limitation concerns the individualization of ethical responsibility. Many AI ethics frameworks implicitly assume that informed users\u0026mdash;clinicians or patients\u0026mdash;can appropriately interpret, contextualize, and contest AI-generated information. The adoption-phase evidence synthesized in this study challenges this assumption. In routine care, time pressure, cognitive load, and asymmetries in expertise significantly constrain meaningful oversight. As LLMs mediate clinical communication, ethical responsibility cannot plausibly rest on individual vigilance alone. Instead, responsibility becomes redistributed across clinicians, institutions, and system designers\u0026mdash;often without explicit acknowledgment or governance. This under-specification of responsibility constitutes a structural ethical vulnerability rather than a remediable implementation flaw.\u003c/p\u003e \u003cp\u003eFinally, prevailing frameworks tend to treat ethical evaluation as a static, pre-deployment exercise, underestimating the dynamic nature of adoption. Ethical risks evolve as patterns of reliance stabilize, user behavior adapts, and organizational norms shift over time. Frameworks that conceptualize ethics as a checklist applied prior to deployment are therefore ill-equipped to detect or respond to emerging harms that arise only through sustained use[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Taken together, these limitations suggest that existing AI ethics frameworks are not merely incomplete, but misaligned with the ethical realities of routine LLM adoption in health care[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eImplications for ethical governance and routine clinical practice\u003c/h2\u003e \u003cp\u003eIf ethical risk emerges primarily during adoption rather than experimentation, ethical responses must likewise shift in focus. First, trust in LLM-supported health-care communication must be institutionally anchored rather than individually negotiated. In practice, users rarely evaluate LLMs solely on technical grounds; instead, they rely on signals of organizational endorsement, accountability, and available recourse. Health-care institutions therefore carry a positive ethical obligation to define how LLM outputs are generated, validated, and monitored within clinical and informational workflows, and to clarify responsibility when AI-mediated information influences communication or decision-making.\u003c/p\u003e \u003cp\u003eSecond, ethical integration of LLMs requires role-sensitive design and governance. Clinicians and patients interact with LLMs under fundamentally different conditions of authority, expertise, and exposure to harm. For clinicians, ethically appropriate systems must support professional accountability, auditability, and calibrated reliance without increasing cognitive burden. For patients and caregivers, ethical use requires plain-language communication, explicit signaling of uncertainty, and safeguards against over-reliance. Treating these roles as ethically interchangeable risks obscuring asymmetries in vulnerability and undermining trust on both sides of the clinical relationship.\u003c/p\u003e \u003cp\u003eThird, equity cannot be treated as a secondary or downstream benefit of technological diffusion. As LLMs increasingly mediate access to medical information, disparities in education, digital capacity, and regional resources translate directly into unequal ethical outcomes. Ethical adoption therefore demands proactive institutional measures, including literacy-sensitive interfaces, adaptive explanatory strategies, and targeted support for underserved populations. Without such measures, LLM integration risks amplifying existing inequities under the guise of informational democratization.\u003c/p\u003e \u003cp\u003eFinally, ethical oversight must be understood as a continuous governance process rather than an episodic intervention. Adoption is not a single event but an evolving trajectory shaped by changing patterns of reliance and organizational practice. Health-care systems must therefore establish mechanisms for ongoing evaluation, feedback, and revision of LLM deployment, recognizing that ethical risks may emerge, intensify, or transform over time. Together, these considerations point toward an ethics of adoption grounded in institutional responsibility, inclusivity, and sustained oversight rather than isolated technical safeguards.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eLarge language models are no longer confined to experimental settings; they are increasingly shaping how medical information is accessed, interpreted, and acted upon in everyday health care. This transition from experimentation to routine adoption constitutes not merely a technological milestone, but a normatively consequential ethical shift. This Article demonstrates that the most pressing ethical challenges associated with LLMs arise less from their technical capabilities than from the ways in which they become embedded within health-care systems, institutional workflows, and social relationships of care.\u003c/p\u003e \u003cp\u003eOur analysis shows that routine adoption systematically reconfigures trust, redistributes responsibility, and exposes asymmetrical vulnerability across user groups. These ethical tensions cannot be adequately addressed through model-level improvements alone. In the absence of institutional accountability, role-sensitive design, and equity-oriented governance, LLM integration risks producing ethically problematic outcomes\u0026mdash;such as misplaced trust, responsibility gaps, and the amplification of existing disparities\u0026mdash;even when technical performance is high. Conversely, when embedded within robust governance structures, LLMs can support clearer communication, enhance understanding, and expand access to health information without undermining professional responsibility or patient protection.\u003c/p\u003e \u003cp\u003eThe ethical task ahead is therefore not to restrain innovation, but to take adoption seriously as an ongoing ethical process rather than a one-time technical achievement. Health-care institutions, researchers, and policymakers must recognize that ethical responsibility does not end at deployment, but persists as patterns of reliance stabilize and evolve over time. By shifting ethical attention from experimental validation toward system-level governance and longitudinal oversight, health-care systems can better ensure that the expanding use of LLMs advances not only efficiency and scale, but also fairness, accountability, and patient trust.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData sharing is not applicable to this article as no datasets were generated or analyzed during the current study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the Talent Fund of Guizhou Provincial People\u0026rsquo;s Hospital (Grant No. 2024-16) and the Science and Technology Fund Project of the Guizhou Provincial Health Commission (Grant No. 2026GZWJKJXM0317). The funding sources had no role in the study design, analysis, interpretation of data, writing of the manuscript, or the decision to submit the manuscript for publication.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026rsquo; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eXY, YW, and JH conceptualized and designed the study. XY led the methodological development and formal ethical analysis. XY, YL, YJL, DL, LY, BZ, and CX contributed to data curation and analysis of relevant empirical literature and adoption studies. XY, YL, YJL, DL, LY, BZ, and CX drafted the initial manuscript. YW and JH critically reviewed and revised the manuscript for important intellectual content and provided overall supervision. YW and JH were responsible for project administration and funding acquisition. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors used the generative AI tool GPT-5 (OpenAI, San Francisco, CA, USA) for language refinement, grammar correction, and consistency editing of the English manuscript. The tool was not used for conceptualization, ethical analysis, interpretation, or reference generation. All scientific content, analysis, and conclusions were developed, verified, and approved by the authors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026rsquo; information\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eMenezes MCS, Hoffmann AF, Tan ALM, Nalbandyan M, Omenn GS, Mazzotti DR, et al. The potential of generative pre-trained transformer 4 (GPT-4) to analyse medical notes in three different languages: a retrospective model-evaluation study. Lancet Digit Health. 2025;7(1):e35\u0026ndash;43. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S2589-7500(24)00246-2\u003c/span\u003e\u003cspan address=\"10.1016/S2589-7500(24)00246-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSun Z, Ong H, Kennedy P, Tang L, Chen S, Elias J, et al. Evaluating GPT4 on Impressions Generation in Radiology Reports. Radiology. 2023;307(5):e231259. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1148/radiol.231259\u003c/span\u003e\u003cspan address=\"10.1148/radiol.231259\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThirunavukarasu AJ, Ting DSJ, Elangovan K, Gutierrez L, Tan TF, Ting DSW. Large language models in medicine. Nat Med. 2023;29(8):1930\u0026ndash;40. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41591-023-02448-8\u003c/span\u003e\u003cspan address=\"10.1038/s41591-023-02448-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSinghal K, Azizi S, Tu T, Mahdavi SS, Wei J, Chung HW, et al. Large language models encode clinical knowledge. Nature. 2023;620(7972):172\u0026ndash;80. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41586-023-06291-2\u003c/span\u003e\u003cspan address=\"10.1038/s41586-023-06291-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang X, Xiao Y, Liu D, Deng H, Huang J, Zhou Y, et al. Cross language transformation of free text into structured lobectomy surgical records from a multi center study. Sci Rep. 2025;15(1):15417. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-025-97500-7\u003c/span\u003e\u003cspan address=\"10.1038/s41598-025-97500-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang X, Zhang Y, Jiang J, Chen Z, Bai R, Yuan Z, et al. Harnessing GPT-4 for automated error detection in pathology reports: Implications for oncology diagnostics. Digit Health. 2025;11:20552076251346703. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1177/20552076251346703\u003c/span\u003e\u003cspan address=\"10.1177/20552076251346703\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang X, Xiao Y, Liu D, Zhang Y, Deng H, Huang J, et al. Enhancing doctor-patient communication using large language models for pathology report interpretation. BMC Med Inf Decis Mak. 2025;25(1):36. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12911-024-02838-z\u003c/span\u003e\u003cspan address=\"10.1186/s12911-024-02838-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang X, Xiao Y, Liu D, Shi H, Deng H, Huang J, et al. Enhancing Physician-Patient Communication in Oncology Using GPT-4 Through Simplified Radiology Reports: Multicenter Quantitative Study. J Med Internet Res. 2025;27:e63786. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2196/63786\u003c/span\u003e\u003cspan address=\"10.2196/63786\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang X, Xiao Y, Liu D, Deng H, Huang J, Zhou Y, et al. Factors Influencing Adoption of Large Language Models in Health Care: Multicenter Cross-Sectional Mixed Methods Observational Study. J Med Internet Res. 2025;27:e84918. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2196/84918\u003c/span\u003e\u003cspan address=\"10.2196/84918\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen J, Liu Y, Liu P, Zhao Y, Zuo Y, Duan H. Adoption of Large Language Model AI Tools in Everyday Tasks: Multisite Cross-Sectional Qualitative Study of Chinese Hospital Administrators. J Med Internet Res. 2025;27:e70789. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2196/70789\u003c/span\u003e\u003cspan address=\"10.2196/70789\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDennst\u0026auml;dt F, Schmerder M, Riggenbach E, Mose L, Bryjova K, Bachmann N, et al. Comparative Evaluation of a Medical Large Language Model in Answering Real-World Radiation Oncology Questions: Multicenter Observational Study. J Med Internet Res. 2025;27:e69752. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2196/69752\u003c/span\u003e\u003cspan address=\"10.2196/69752\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShin HS, Williams H, Braykov N, Jahan A, Meller J, Orenstein EW. The Influence of Artificial Intelligence Scribes on Clinician Experience and Efficiency among Pediatric Subspecialists: A Rapid, Randomized Quality Improvement Trial. Appl Clin Inf. 2025;16(4):1041\u0026ndash;52. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1055/a-2657-8087\u003c/span\u003e\u003cspan address=\"10.1055/a-2657-8087\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHanna MG, Pantanowitz L, Jackson B, Palmer O, Visweswaran S, Pantanowitz J, et al. Ethical and Bias Considerations in Artificial Intelligence/Machine Learning. Mod Pathol. 2025;38(3):100686. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.modpat.2024.100686\u003c/span\u003e\u003cspan address=\"10.1016/j.modpat.2024.100686\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTorkamaan H, Steinert S, Pera MS, Kudina O, Freire SK, Verma H, et al. Challenges and future directions for integration of large language models into socio-technical systems. Behav Inform Technol. 2024;1\u0026ndash;20. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1080/0144929X.2024.2431068\u003c/span\u003e\u003cspan address=\"10.1080/0144929X.2024.2431068\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDhagarra D, Goswami M, Kumar G. Impact of Trust and Privacy Concerns on Technology Acceptance in Healthcare: An Indian Perspective. Int J Med Inf. 2020;141:104164. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.ijmedinf.2020.104164\u003c/span\u003e\u003cspan address=\"10.1016/j.ijmedinf.2020.104164\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKaur D, Uslu S, Rittichier KJ, Durresi A. Trustworthy Artificial Intelligence: A Review. ACM Comput Surv. 2022;55(2).):Article 39. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1145/3491209\u003c/span\u003e\u003cspan address=\"10.1145/3491209\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSinghal A, Neveditsin N, Tanveer H, Mago V, Toward Fairness. Accountability, Transparency, and Ethics in AI for Social Media and Health Care: Scoping Review. JMIR Med Inf. 2024;12:e50048. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2196/50048\u003c/span\u003e\u003cspan address=\"10.2196/50048\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVollmer S, Mateen BA, Bohner G, Kir\u0026aacute;ly FJ, Ghani R, Jonsson P, et al. Machine learning and artificial intelligence research for patient benefit: 20 critical questions on transparency, replicability, ethics, and effectiveness. BMJ. 2020;368:l6927. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1136/bmj.l6927\u003c/span\u003e\u003cspan address=\"10.1136/bmj.l6927\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMassenon R, Gambo I, Khan JA, Agbonkhese C, Alwadain A. My AI is Lying to Me: User-reported LLM hallucinations in AI mobile apps reviews. Sci Rep. 2025;15(1):30397. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-025-15416-8\u003c/span\u003e\u003cspan address=\"10.1038/s41598-025-15416-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVisser R, Peters TM, Scharlau I, Hammer B. Trust, distrust, and appropriate reliance in (X)AI: A conceptual clarification of user trust and survey of its empirical evaluation. Cogn Syst Res. 2025;91:101357. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.cogsys.2025.101357\u003c/span\u003e\u003cspan address=\"10.1016/j.cogsys.2025.101357\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim SH, Wihl J, Schramm S, Berberich C, Rosenkranz E, Schmitzer L, et al. Human-AI collaboration in large language model-assisted brain MRI differential diagnosis: a usability study. Eur Radiol. 2025;35(9):5252\u0026ndash;63. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s00330-025-11484-6\u003c/span\u003e\u003cspan address=\"10.1007/s00330-025-11484-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEbnali Harari R, Altaweel A, Ahram T, Keehner M, Shokoohi H. A randomized controlled trial on evaluating clinician-supervised generative AI for decision support. Int J Med Inf. 2025;195:105701. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.ijmedinf.2024.105701\u003c/span\u003e\u003cspan address=\"10.1016/j.ijmedinf.2024.105701\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAllen LN, Lin J, Segal BM, Ndlovu K, Bilardi D, Pettigrew LM. Artificial intelligence in primary care: frameworks, challenges, and guardrails. Lancet Prim Care. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.lanprc.2025.100079\u003c/span\u003e\u003cspan address=\"10.1016/j.lanprc.2025.100079\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHeston TF, Lu M, Gillette J. Ethical Hazards of Large Language Models in Primary Care: A Clinician-Focused Update. In: Heston TF, editor. Medical Ethics - Navigating Complex Decisions in Contemporary Healthcare. London: IntechOpen; 2025.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRyan K, Kasun M, Roberts LW, Kim JP. Information, collaboration, regulation: Physician and AI researcher views on ethical considerations in clinical AI integration. Big Data Soc. 2025;12(2):20539517251343853. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1177/20539517251343853\u003c/span\u003e\u003cspan address=\"10.1177/20539517251343853\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYıldırım C, Aykut A, G\u0026uuml;nsoy E, \u0026Ouml;nc\u0026uuml;l MV. Evaluating GPT-4o for emergency disposition of complex respiratory cases with pulmonology consultation: a diagnostic accuracy study. Scand J Trauma Resusc Emerg Med. 2025;33(1):159. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s13049-025-01475-3\u003c/span\u003e\u003cspan address=\"10.1186/s13049-025-01475-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee S, Jung S, Park JH, Cho H, Moon S, Ahn S. Performance of ChatGPT, Gemini and DeepSeek for non-critical triage support using real-world conversations in emergency department. BMC Emerg Med. 2025;25(1):176. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12873-025-01337-2\u003c/span\u003e\u003cspan address=\"10.1186/s12873-025-01337-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAzevedo CB, Martinho AS, Braga I, Nogueira-Silva C, Barroso C, Correia-Pinto J. ChatGPT-4o in Enhancing Informed Consent in Pediatric Surgical Practice. J Pediatr Surg. 2025;60(9):162413. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jpedsurg.2025.162413\u003c/span\u003e\u003cspan address=\"10.1016/j.jpedsurg.2025.162413\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eUmman V, Tosun B, Uygur A, Emre S. Evaluation of the Effectiveness, Safety, and Patient Satisfaction of Artificial Intelligence-Based Patient Education and Counseling for Both Recipients and Donors in the Preoperative and Postoperative Phases of Organ Transplantation. Transplant Proc. 2025;57(9):1832-9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.transproceed.2025.07.001\u003c/span\u003e\u003cspan address=\"10.1016/j.transproceed.2025.07.001\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNouis SC, Uren V, Jariwala S. Evaluating accountability, transparency, and bias in AI-assisted healthcare decision- making: a qualitative study of healthcare professionals' perspectives in the UK. BMC Med Ethics. 2025;26(1):89. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12910-025-01243-z\u003c/span\u003e\u003cspan address=\"10.1186/s12910-025-01243-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFink A, Nattenm\u0026uuml;ller J, Rau S, Rau A, Tran H, Bamberg F, et al. Retrieval-augmented generation improves precision and trust of a GPT-4 model for emergency radiology diagnosis and classification: a proof-of-concept study. Eur Radiol. 2025;35(8):5091\u0026ndash;8. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s00330-025-11445-z\u003c/span\u003e\u003cspan address=\"10.1007/s00330-025-11445-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFukui Y, Kawata Y, Kobashi K, Nagatani Y, Iguchi H. Evaluation of a retrieval-augmented generation system using a Japanese Institutional Nuclear Medicine Manual and large language model-automated scoring. Radiol Phys Technol. 2025;18(3):861\u0026ndash;76. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s12194-025-00941-y\u003c/span\u003e\u003cspan address=\"10.1007/s12194-025-00941-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDang Q, Li G. Unveiling trust in AI: the interplay of antecedents, consequences, and cultural dynamics. AI Soc. 2025. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s00146-025-02477-6\u003c/span\u003e\u003cspan address=\"10.1007/s00146-025-02477-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eUrbina JT, Vu PD, Nguyen MV. Disability Ethics and Education in the Age of Artificial Intelligence: Identifying Ability Bias in ChatGPT and Gemini. Arch Phys Med Rehabil. 2025;106(1):14\u0026ndash;9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.apmr.2024.08.014\u003c/span\u003e\u003cspan address=\"10.1016/j.apmr.2024.08.014\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJi Y, Ma W, Sivarajkumar S, Zhang H, Sadhu EM, Li Z, et al. Mitigating the risk of health inequity exacerbated by large language models. npj Digit Med. 2025;8(1):246. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41746-025-01576-4\u003c/span\u003e\u003cspan address=\"10.1038/s41746-025-01576-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNaghdi M, Cao P, Essers R, Heijligers M, Paulussen ADC, van der Lugt A, et al. Artificial intelligence-simplified information to advance reproductive genetic literacy and health equity. Hum Reprod. 2025;40(9):1681\u0026ndash;8. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/humrep/deaf135\u003c/span\u003e\u003cspan address=\"10.1093/humrep/deaf135\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAhmed H. Large language models for clinical trials in the Global South: opportunities and ethical challenges. AI Ethics. 2025;6(1):76. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s43681-025-00943-x\u003c/span\u003e\u003cspan address=\"10.1007/s43681-025-00943-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRodler S, Cei F, Ganjavi C, Checcucci E, De Backer P, Rivero Belenchon I, et al. GPT-4 generates accurate and readable patient education materials aligned with current oncological guidelines: A randomized assessment. PLoS ONE. 2025;20(6):e0324175. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pone.0324175\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0324175\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCampellone TR, Flom M, Montgomery RM, Bullard L, Pirner MC, Pavez A, et al. Safety and User Experience of a Generative Artificial Intelligence Digital Mental Health Intervention: Exploratory Randomized Controlled Trial. J Med Internet Res. 2025;27:e67365. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2196/67365\u003c/span\u003e\u003cspan address=\"10.2196/67365\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Y, Calle P, Vadakekut M, Rubin D, Nagykaldi Z, Doescher M, et al. AI-Enabled Personalized Smoking Cessation Intervention With the Aipaca Chatbot: Mixed Methods Feasibility Study. JMIR Form Res. 2025;9:e73319. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2196/73319\u003c/span\u003e\u003cspan address=\"10.2196/73319\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\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":"Large language models, Medical ethics, Ethical governance, Health-care communication, Technology adoption, Equity and trust, Responsibility","lastPublishedDoi":"10.21203/rs.3.rs-8553262/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8553262/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eLarge language models (LLMs) are increasingly embedded in routine health-care communication, extending beyond experimental evaluation to real-world use by clinicians, patients, and caregivers. This transition from experimentation to adoption represents a normatively consequential ethical shift. While early research has focused primarily on technical performance, emerging ethical challenges arise less from model capabilities than from how LLMs become integrated into clinical workflows, institutional arrangements, and relationships of care.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eWe conducted a normative ethical analysis informed by empirical evidence from recent real-world studies of LLM adoption in health-care settings. Drawing on adoption-oriented empirical research, we examined how the transition from experimental use to routine reliance reshapes ethical conditions related to trust, responsibility allocation, and equity across different user groups.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe analysis identifies a systematic shift in ethical risk from model-level concerns\u0026mdash;such as accuracy, validation, and bias mitigation\u0026mdash;toward system-level dynamics that emerge during routine adoption. Adoption alters how medical information is mediated, redistributes responsibility among clinicians, institutions, and patients, and exposes asymmetrical vulnerability linked to digital literacy, educational background, and regional context. Rather than uniformly democratizing access to medical knowledge, routine LLM use may function as a capability amplifier, magnifying existing inequities in the absence of institutional safeguards.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eEthical challenges arising from routine LLM adoption in health care cannot be adequately addressed through model-level or pre-deployment ethics alone. Addressing these challenges requires adoption-sensitive governance approaches that recognize reliance as an ongoing ethical process. Institutional accountability, role-sensitive design, and continuous oversight are ethically necessary to ensure that LLM integration enhances health-care communication and access without eroding trust or exacerbating existing disparities.\u003c/p\u003e","manuscriptTitle":"From experimentation to adoption: a normative ethical analysis of large language models in health care","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-23 00:36:41","doi":"10.21203/rs.3.rs-8553262/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":"aa6c6ae1-2ce3-41c5-bdbe-39caa03b6fb6","owner":[],"postedDate":"January 23rd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-02-11T05:25:21+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-23 00:36:41","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8553262","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8553262","identity":"rs-8553262","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}