This study strictly followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines (Checklist 1) [18] and the ENTREQ (Enhancing Transparency in Reporting the Synthesis of Qualitative Research; Checklist 2) [19]. This review aimed to explore patients’ values, attitudes, and experiences regarding the application of AI in health care. The Cochrane Handbook recognizes qualitative evidence synthesis as the appropriate methodology for such questions [20]. As all 25 included studies used qualitative designs yielding narrative findings rather than quantitative effect estimates amenable to statistical pooling [21], statistical meta-analysis was not applicable to this review. We therefore used the Joanna Briggs Institute meta-aggregation approach [22]. The review was prospectively registered in PROSPERO (CRD420251156502). The registered protocol was subsequently updated to clarify the qualitative study design and the use of JBI-QARI (Joanna Briggs Institute Qualitative Assessment and Review Instrument) for quality assessment. In addition, the following methodological enhancements were made during the conduct of this review: (1) CINAHL was added as a fifth database to improve coverage of nursing and allied health literature, (2) the GRADE-CERQual (Confidence in the Evidence from Reviews of Qualitative Research) approach was adopted to assess confidence in the synthesized findings, and (3) social ecological theory (SET) was used as a theoretical lens to interpret and discuss the findings.

This study’s search strategy followed the PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses literature search extension) guideline (Checklist 3) [23]. The initial search was conducted on September 28, 2025. Following iterative refinement of the search strategy, updated searches were performed on January 4, 2026, and March 1, 2026, to capture the most recent literature. Five databases were searched from inception to the date of each search: PubMed (via National Library of Medicine), Embase (via Embase.com), Web of Science (via Clarivate, encompassing the Core Collection, KCI-Korean Journal Database, MEDLINE, ProQuest Dissertations & Theses Citation Index, SciELO Citation Index, and the Grants Index), CINAHL (via EBSCOhost), and Scopus (via Scopus.com); all other databases were searched independently through their respective platforms. The detailed search strategies for all databases are provided in Multimedia Appendix 1. A 3-step search strategy was used. First, an initial search was conducted in PubMed, and the titles, abstracts, and index terms of relevant records were analyzed to identify key search terms. Second, a comprehensive search using all identified keywords and index terms was undertaken across all databases. Third, the reference lists of included studies were hand-searched to identify any additional relevant studies. The search strategy was developed de novo for this review, and its effectiveness was validated by confirming the retrieval of known relevant studies. No formal peer review of the search strategy was conducted using standardized appraisal tools. This review focused on published qualitative research; clinical trial registries were not searched. Beyond the systematic database searches and reference list screening, no supplementary search methods were used, such as contacting authors, browsing conference proceedings, or setting up citation alerts. Only studies published in English or Chinese were included. The PubMed search strategy is presented below:

This review focused on patient concerns arising from the use of AI in clinical practice. The inclusion and exclusion criteria are presented in Table 1. A total of 19,090 records were retrieved and imported into EndNote (version 21; Clarivate Analytics). Through both automated and manual deduplication in EndNote 21, a total of 7132 (37.4%) duplicate records were identified, leaving 11,958 (62.6%) records for assessment based on the relevance of titles and abstracts. At the title and abstract screening stage, 11,874 (99.3%) records were excluded for the following reasons: not addressing AI applications in clinical health care settings (n=5818, 49%), not involving patients as participants (n=3871, 32.6%), ineligible study design (n=2030, 17.1%), and not published in English or Chinese (n=155, 1.3%). A total of 84 (0.7%) records met the selection criteria, and their full texts were retrieved for further evaluation. Following full-text assessment, 25 studies met the criteria for quality appraisal. Two authors independently conducted the screening process, and any disagreements regarding inclusion were resolved through consultation with a third author.

The JBI-QARI was used to evaluate the methodological rigor of each published study [24]. Questions answered “yes” received 1 point, and studies scoring 5 points or lower were deemed low quality and excluded from the synthesis. Two reviewers independently conducted rigorous assessments of the selected research reviews. Disagreements were resolved through discussion or consultation with a third reviewer within the team. Table 2 presents the quality assessment results for the studies included in this review.

Data were extracted from the studies included in this review using the JBI-QARI standardized data extraction tool. The first author extracted data from these 25 studies, including the first author and publication year, study population and sample size, research topic, study type, and primary outcomes, as detailed in Table 3. This study used a meta-synthesis approach [22]. Two researchers repeatedly read and interpreted the original studies, analyzed and interpreted the implications of the findings, grouped similar results into new categories, and then synthesized these categories into integrated outcomes to form new perspectives or interpretations. When disagreements arose between the 2 coders, a third coder was consulted.

The confidence in each synthesized finding was assessed using the GRADE-CERQual approach [50]. CERQual assesses confidence based on four components: (1) methodological limitations of the included studies, informed by the JBI-QARI critical appraisal results; (2) coherence of the finding across contributing studies; (3) adequacy of data supporting the finding, considering both the number of studies and the richness of data; and (4) relevance, defined as the extent to which the contexts of contributing studies are applicable to the review question. Each finding was assigned an overall confidence level of high, moderate, low, or very low. Two reviewers independently assessed each component and resolved disagreements through discussion. The CERQual assessment results are presented in Table 4. Heterogeneity across included studies was explored narratively by examining differences in AI application types, patient populations, and geographic contexts, as reported in Table 3 and discussed in the Limitations section.

This paper provides a systematic review and qualitative synthesis of ethical issues surrounding the application of AI in health care. A total of 25 qualitative studies involving 528 patients were included. Details of the search and screening process are presented in the PRISMA flow diagram (Figure 1). The findings were categorized into 14 themes and further consolidated into 6 key themes (Figure 2), elaborated as follows.

Patients believe that while AI can provide technical support, it cannot replace the human care and ethical judgment physicians offer in treatment. AI cannot comprehend patients’ emotional needs or life contexts, factors crucial to treatment decisions. Patients emphasized that AI should only serve as an auxiliary tool, unable to substitute for physicians’ roles in complex medical decisions—particularly regarding emotional support and personalized treatment.

This study aimed to conduct an in-depth analysis of patients’ ethical concerns regarding AI medical applications, explore their impact on patient acceptance and trust, and use SET to reveal multilevel tension mechanisms. SET, which is based on Bronfenbrenner’s ecological systems theory [51], was later improved by researchers such as McLeroy and has become a well-known way to study public health and health behavior [52,53]. The theory’s core assumption posits that individual behavior does not exist in isolation but is embedded within a multilevel, interacting environmental system. The microlevel focuses on individual knowledge, attitudes, and psychological characteristics; the mesolevel involves interpersonal relationships and organizational environments; and the macrolevel encompasses sociocultural norms and policy systems [52,54]. These 3 levels form a complex ecosystem through dynamic bidirectional influences, where changes in any one level may propagate to others.

Through a meta-synthesis of 25 qualitative studies, we identified 6 core themes: privacy and data security, technological reliability, impacts on physician-patient relationships, trust and accountability, ethical challenges and health care equity, and future perspectives. Socioecological analysis reveals that these concerns create mutually reinforcing “ecological imbalances” across micro (technological cognitive biases and data control anxieties), meso (broken physician-patient trust and institutional accountability gaps), and macro (health inequities and lagging ethical standards) levels, thereby hindering patient acceptance and trust in AI technologies. Using this framework, the following sections will examine these findings in detail.

At the individual level, patients’ concerns about AI technology primarily manifest as cognitive uncertainty about the technology and anxiety over data control. SET emphasizes the interaction between individual behavior and multilevel environments, where individuals’ cognitive and emotional attitudes are influenced not only by internal factors but also closely tied to their social and cultural contexts [55]. During the application of AI medical technologies, individuals’ cognitive biases about the technology are often accompanied by intense concerns about its transparency and controllability, which can lead to hesitance in adopting these technologies for their health care needs.

Research indicates that patients’ understanding of AI technology is closely linked to their acceptance of it [56]. However, the complexity of AI technology and the black box nature of algorithms make it difficult for patients to comprehend its decision-making processes. This uncertainty directly leads to patients questioning the technology’s accuracy [57]. For instance, patients worry that AI cannot account for their individualized health variations—such as differences in underlying conditions or lifestyle habits—potentially compromising the effectiveness of medical interventions [58]. Furthermore, patients’ doubts about AI accuracy have extended from the algorithmic level to the data level. Drawing from experiences reviewing their own electronic health records, they have identified numerous errors in the documentation. Consequently, they fear that even well-designed algorithms trained on such flawed data may struggle to produce reliable outputs. This technological cognitive bias does not exist in isolation; it is intertwined with individuals’ anxieties over data control. Patients’ concerns about data control cannot be reduced to instrumental worries about privacy breaches but should be understood as existential threats to their self-integrity and narrative sovereignty [59]. Patients’ anxiety over data control extends beyond privacy leaks to a profound fear of losing self-determination [60]. Medical data encapsulate patients’ life experiences, health statuses, and identity information. When processed by algorithms for unforeseen purposes, patients lose not only informational control but also dominion over their own health narratives [61-63].

SET posits that individual behavior is shaped by multiple interrelated factors [55]. At the microlevel, patients’ technological cognitive biases and data control anxieties do not exist in isolation but form a vicious cycle through the “cognition-anxiety” interaction [64]. Distrust of AI technology not only heightens patients’ concerns about data security but also deepens their skepticism regarding technological transparency and controllability [65]. This vicious cycle may lead to patient resistance toward AI technology, subsequently affecting their overall attitude toward health care services. Similar perspectives are reflected in other studies. Research indicates that uncertainty about information technology and privacy concerns often amplify individual resistance, thereby influencing their acceptance of technology [66].

At the mesolevel, patient concerns primarily manifest as fractures in physician-patient trust and inadequacies in health care organizational governance. Interpersonal relationships and organizational culture often constrain individual behavioral changes, according to SET [55]. Patients’ apprehensions toward AI medical technologies stem not only from individual perceptions of the technology but are also closely intertwined with the quality of physician-patient relationships and the accountability of health care organizations.

Extensive research indicates that emotional empathy and interpersonal interaction between physicians and patients are crucial for building trust [67-69]. However, the introduction of AI has undermined this foundation to some extent. Studies reveal that when physicians overrely on AI technology, patients experience reduced interaction time and emotional support from their physicians, leading to diminished trust [70]. The core of the physician-patient relationship lies in physicians viewing patients as whole individuals, not machines requiring repair [71]. This relationship is based on listening, understanding, and being there for each other. Physicians know how much their patients are hurting and help them by talking to them [72-74]. AI intervention threatens the quality of this interpersonal engagement. When physicians’ attention shifts to screens, when diagnoses rely on algorithmic outputs, and when communication is replaced by standardized processes, patients transform from “people receiving care” into “objects undergoing testing” [70,75].

Conversely, the absence of accountability mechanisms within health care organizations is another significant source of patient concern. Existing research indicates that many institutions lack clear AI liability frameworks, leaving patients unable to identify responsible parties when AI systems malfunction [76]. This lack of accountability not only makes patients less trusting of AI technology but also makes health care organizations look less credible [77]. SET emphasizes that deficiencies in internal organizational governance structures amplify individual-level concerns by eroding patients’ trust in the health care system [78]. The absence of accountability in AI decision-making within health care organizations leaves patients without effective avenues for redress when encountering medical issues, thereby creating a mutually reinforcing negative cycle of “accountability-trust” [79]. Similar “accountability vacuum” issues have been explored in other studies, with research indicating that patients’ trust in medical decisions is often severely compromised when health care organizations lack clear responsibility frameworks [80].

At the macrolevel, patient concerns regarding AI technology primarily center on health care equity and the lag in ethical regulations. SET asserts that sociocultural inequities and the lack of ethical standards directly impact individual behavioral choices [81]. During the implementation of AI health care technologies, widespread adoption encounters obstacles arising from socioeconomic disparities. Low-income groups and patients with limited technological literacy often struggle to access AI medical services equitably. This gap exacerbates health care inequalities, fueling patient resistance toward AI technologies [82,83].

Not all socioeconomic strata equally benefit from the application of AI health care technologies, especially in regions with slower economic development or scarce resources [84]. High-end medical facilities and advanced technological resources are concentrated in major cities and economically developed regions, while low-income communities and remote areas still face significant gaps in AI health care adoption [85]. Furthermore, the high technical support and maintenance requirements of AI health care technologies make them unaffordable for many resource-constrained medical institutions, further exacerbating the unequal distribution of health care resources [86]. In this context, the “technological access barriers” experienced by patients are not merely technical difficulties but deep-seated social problems stemming from unequal socioeconomic structures and resource allocation [87].

Although the World Health Organization (WHO) has issued global AI ethics principles (WHO, 2021), implementation and regulatory rigor vary significantly across countries [88]. The lag in ethical standards manifests not only at the policy level but also creates gaps in practical implementation. For instance, some countries may lack cross-regional ethical collaboration mechanisms, leading certain health care institutions to prioritize AI services for economically advantaged groups over low-income populations due to cost considerations [89]. Such practices exacerbate health inequalities, deepening patients’ concerns that AI serves only select groups rather than benefiting the broader public [90].

According to SET, concerns at the microlevel, mesolevel, and macrolevel do not exist in isolation but form mutually reinforcing chain reactions through “risk perception transmission” [55]. Research indicates that individual cognition and anxiety at the microlevel generate reverse effects at interpersonal and organizational levels, subsequently impacting broader social structures [91]. At the microlevel, individuals’ cognitive biases regarding AI technology and anxieties over data control rights are transmitted to the mesolevel through interactions between individuals and health care organizations, increasing communication pressures for these organizations when deploying AI technologies [92,93]. Patients’ concerns about data privacy and security influence individual behavioral decisions and prompt health care institutions to reevaluate the boundaries of AI applications, thereby driving societal-level attention to AI ethical norms [94].

The absence of accountability mechanisms for medical organizations at the mesolevel exacerbates microlevel cognitive biases and anxieties, leading to increased distrust among patients and further complicating the integration of AI technologies in health care [93,95]. Extensive research indicates that the breakdown of social support networks directly impacts individual health decisions and attitudes, with trust deficits further intensifying emotional distress and technological apprehension [96-98]. The breakdown of physician-patient trust not only reduces patient acceptance of AI technology but also leads patients to rely more on personal emotional judgments when facing medical decisions, overlooking the potential of AI technology [99].

Social equity issues at the macrolevel further permeate the microlevel, particularly the rejection of AI health care services by low-income groups. This affects their acceptance of AI technology and reinforces fears of technological displacement through sociocultural perceptions [17,100]. SET indicates that sociocultural beliefs not only shape patients’ perceptions of AI technology through individual behavior but also influence organizational behavior via mesolevel cultural diffusion, thereby amplifying implementation barriers in unequal societies [101,102].

Reflecting on the findings of this study, patients’ concerns are far from unfounded. Across the 25 studies included in this review, patients from diverse clinical contexts and cultural backgrounds consistently expressed anxieties about data security, skepticism toward algorithmic opacity, and questions about accountability attribution. These concerns point to objective limitations of AI medical technologies at their current stage of development and institutional gaps that remain unaddressed. Recent studies have independently corroborated the reality of these issues from various perspectives: data bias and privacy risks in clinical AI systems have been extensively documented [103], the erosion of physician-patient trust caused by algorithmic inexplicability has attracted sustained attention [104], and accountability attribution in AI-assisted medical decision-making still lacks clear legal delineation [105]. The WHO, in its 2021 guidance, also identified transparency, accountability, and equity as core principles for the ethical governance of AI in health care [106]. In other words, the concerns articulated by patients based on their lived experiences correspond precisely with the risks identified through systematic analyses by the academic community and international organizations. This implies that the goal of policy intervention should not be to “correct patient misconceptions,” but rather to substantively address these well-founded and evidence-based concerns. The following recommendations are organized across the microlevel, mesolevel, and macrolevel while acknowledging the implementation challenges inherent in each pathway.

Given that the field of Explainable Artificial Intelligence remains in its developmental stages, fully visualizing algorithmic decision-making logic is not realistic in the short term; an incremental transparency strategy should therefore be adopted [107]. Specifically, drawing on the access rights for data subjects under the European Union’s General Data Protection Regulation [108], health care institutions should be required to establish data access logging systems that enable patients to query the access records and intended purposes of their health data [109]. To address patients’ concerns regarding the quality of training data, accessible mechanisms for medical record review and correction should be established, safeguarding patients’ rights to audit and amend their own medical records [110]. Furthermore, informed consent regarding AI involvement should be front-loaded—patients should be notified before their clinical encounter—and channels for questioning AI judgments and requesting human review should be established [111].

The principle of “human oversight, AI assistance” should be institutionally safeguarded. Rather than imposing rigid communication time benchmarks, mandatory “physician confirmation checkpoints” should be embedded within AI-assisted diagnostic and treatment workflows, ensuring that critical decisions are subject to physician review and explained to patients before implementation [112]. Concurrently, performance evaluation systems should be adjusted to prevent efficiency-driven metrics from encroaching upon the space for physician-patient communication [113]. Regarding accountability attribution, a tiered liability framework should be constructed that differentiates the responsible parties for data errors, algorithmic defects, and clinical misjudgments [114], while acknowledging that existing legal frameworks contain gaps in the attribution of responsibility for AI-driven decisions, necessitating legislative follow-up [115]. Additionally, clinical scenarios in which AI cannot substitute for human practitioners should be explicitly delineated, particularly those highly dependent on emotional support and clinical judgment, such as end-of-life care, mental health treatment, and complex decision-making situations that require empathy and nuanced understanding of patient needs [116].

The WHO has set out global ethical standards for AI in health care, which can be used to help create an international ethical framework. National governments should build upon this foundation to develop regulatory frameworks aligned with their respective health care systems and cultural traditions, rather than pursuing an unrealistic goal of globally uniform standards. At the level of resource allocation, investment should be increased for low-income populations and remote areas, and age-friendly and low-cost AI health care tools should be developed to narrow the gap in technological accessibility [117]. However, it must be acknowledged that the barriers facing these regions extend beyond equipment scarcity to include insufficient technical maintenance capacity and digital health literacy, necessitating complementary capacity building and educational support [118]. Furthermore, equity assessments should be incorporated into the market approval review of AI health care products, requiring developers to submit impact assessment reports for diverse population groups [119].

Finally, it must be acknowledged that structural tensions exist in the advancement of AI in health care: technology developers’ pursuit of algorithmic efficiency and commercial returns may conflict with patients’ safety needs; health care institutions under cost pressures may find it difficult to reconcile these with adequate physician-patient communication. Confronting rather than evading these tensions is a prerequisite for formulating pragmatic policies [120].

This study has the following limitations. First, the 25 included studies were primarily conducted in developed regions, such as North America, Europe, and Australia, with insufficient representation of patients from developing countries. This limits the generalizability of findings to resource-constrained settings. Second, the included studies covered a wide range of AI applications—including imaging diagnostics, clinical decision support, and virtual health assistants—where patient concerns may vary across different AI types. The thematic synthesis in this study may obscure such context-specific differences. Third, as a secondary analysis, the quality of the meta-synthesis depends on the reporting depth of the original studies. Some studies provided insufficient contextual information for patient quotes, hindering more nuanced interpretation. Finally, AI medical technology is undergoing rapid iteration. The literature included in this study reflects patient perceptions within a specific time window. As technological transparency and regulatory frameworks evolve, patient attitudes may change, potentially leading to increased trust in AI medical technology and greater acceptance of its use in health care. Future research should continuously track this dynamic process.

This research, through a systematic synthesis of 25 qualitative studies, identified 6 primary patient concerns regarding AI health care applications: privacy and data security, technological reliability, the impact on physician-patient relationships, trust and accountability, ethics and fairness, and ambivalent attitudes toward future developments. Unlike previous reviews focusing on the general public, this study centers on patients as core stakeholders. It pioneers the application of SET to this field, revealing a “disrupted ecological equilibrium” mechanism that propagates across microlevel, mesolevel, and macrolevel. This provides an explanatory framework—transcending descriptive induction—for understanding the deep-seated reasons behind patient resistance to AI medical technologies. The findings offer direct implications for practice: clinical institutions should establish a “human-led, AI-assisted” diagnostic model; policymakers should accelerate liability legislation and prioritize equitable technology access; and developers should adopt incremental transparency strategies while providing patients with avenues for questioning and review. Future research may explore the following directions: incorporating perspectives from patients in more developing countries and resource-constrained regions to test cross-cultural applicability; conducting comparative studies across different AI application scenarios to explore the context-specific nature of concerns; and using longitudinal designs to track the dynamic evolution of patient attitudes as technology advances and regulatory frameworks mature.