Our systematic review follows the Cochrane Collaboration Handbook [14,15] and reports findings in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 checklist [16]. This review systematically consolidates findings from the past 5 years to address underexplored areas of health care workers’ trust in AI-CDSSs, identifying both enablers and barriers within trust dynamics [14-16]. The Critical Appraisal Skills Programme (CASP) tool [17] was used to assess the quality of included studies and the risk of bias.

We conducted a systematic search of published studies focusing on trust in AI-CDSS between January 1, 2020, and November 30, 2024, guided by PRISMA guidelines with the PICO (population, intervention, comparison, and outcome) framework [14,16]. This study period was chosen to reflect the advancements and increased investment in AI, particularly following the release of generative models such as ChatGPT in the health care sector, especially in the aftermath of the COVID-19 pandemic. Our sources included PubMed, Scopus, and Google Scholar. The search strategy used a combination of English keywords, including “trust” or “acceptance” or “perception” and “artificial intelligence” or “AI” and “decision support systems” or “clinical decision support” or “AI-based decision support” and “healthcare workers” or “clinicians” or “nurses” or “medical professionals” or “healthcare providers.” Publication date filters were applied to include only studies within the specified time frame. Additionally, we used a snowball strategy to identify further sources from the references of relevant full texts. Medical Subject Headings (MeSH) and free-text terms were used to maximize search sensitivity and ensure comprehensive coverage of relevant literature, as described in Multimedia Appendix 1.

We included research articles that explicitly described health care professionals’ trust, acceptance, or reliance on AI-CDSSs, specifically within clinical and primary care settings. Qualitative, quantitative, and mixed-method studies were all considered, provided they explored aspects of trust or acceptance among health care workers. We excluded studies unrelated to the relationship between trust in health care providers and AI-CDSSs. To maintain the scientific credibility of our review, we also excluded non–peer-reviewed articles, editorials, opinion pieces, and other forms of nonresearch literature.

Two researchers (HMT and HAR) initially screened titles and abstracts to determine whether they met the inclusion criteria. After removing duplicates, full texts were reviewed to assess potential exclusion criteria. Any disagreements regarding eligibility criteria were resolved through discussion with another team member. The Mixed Methods Appraisal Tool was used to assess the quality of studies, enabling evaluation across diverse methodological approaches [18]. We extracted key study details, including author, country of data origin, study design, and the type of AI method utilized. Additional information was systematically extracted for each study, focusing on the type of AI application, the role of health care workers, the study setting and location, the specific department or clinical focus, and the type of AI-based decision support system. We also recorded information on trust measurement tools, qualitative questions related to trust, assessed trust factors, trust-related outcomes, levels of trust observed, influencing factors, study limitations, conclusions or recommendations, and funding sources. Additionally, qualitative information, such as quotes, themes, and findings from interviews, focus groups, and open-ended survey responses, was extracted. The quality of the included studies was also evaluated based on the alignment between study objectives and results.

Data synthesis and analysis were conducted to systematically integrate and interpret the findings. Relevant data were organized into an evidence matrix using a standardized template in Google Sheets (Google LLC/Alphabet Inc). Advanced tools for systematic review and data extraction, including Zotero 6 (Corporation for Digital Scholarship), Elicit (Ought), and Rayyan (Qatar Computing Research Institute), were used to screen and analyze abstracts from 27 studies that met the inclusion and exclusion criteria. A comprehensive data charter was developed to summarize the characteristics of the included study [19]. This data charter includes the year of publication, geographic location, study setting, characteristics of CDSSs, methods used to evaluate trust, descriptions of the AI-CDSSs, and evaluation of trust-related factors in AI-CDSSs. Furthermore, we extracted qualitative outcomes and corresponding codes related to trust in AI-CDSSs. The extracted data were organized and analyzed in Microsoft Excel, with recurrent patterns categorized into themes using mind-mapping techniques. Each theme was carefully interpreted and synthesized into enablers and barriers to trust in AI-CDSSs. Based on these themes, we developed actionable recommendations for practical implementation, policy, and further research by triangulating insights from thematic synthesis, extracted quotes, and relevant literature.

The quality of qualitative studies and the qualitative components of mixed-methods studies were assessed using the CASP tool. Each question in the tool was evaluated with 1 of 3 responses: “yes,” “no,” or “cannot determine.” Rather than producing a summative score, the CASP tool provides an overall assessment, categorizing studies as “not valuable,” “semivaluable,” “valuable,” or “very valuable,” as documented in previous literature. These assessments were based on a judgmental approach, in which reviewers evaluated the relevance and contribution of each study to understanding interaction traits within the AI-clinician quality of interaction construct [17]. Quality assessments were conducted independently by 2 reviewers (OAM and LN), with any disagreements resolved through consensus. This process ensured a rigorous and transparent evaluation of study quality.

The article selection process consisted of 2 phases: (1) a review of titles and abstracts and (2) a full-text review. Figure 1 illustrates the study selection process. Initially, 333 records were identified from 3 databases: PubMed (69 records), Scopus (142 records), and Google Scholar (122 records). An independent reviewer screened these records to remove duplicates and articles deemed irrelevant based on titles and abstracts, resulting in 60 records advancing to the next stage. Further screening excluded 20 articles due to unsuitable study designs, leaving 40 for eligibility assessment. Following an in-depth full-text review and final discussions among the reviewers, 13 articles were excluded for lacking a focus on trust in AI-based decision support systems among health care workers. Ultimately, 27 studies were included in the final analysis.

A total of 23 studies were assessed using the CASP checklist for qualitative analysis (Table 1), while 4 studies used other methodologies. Among the 23 studies, the majority (n=19, 83%) received a “Yes” rating for most CASP criteria and were categorized as “valuable” or “very valuable” in the quality assessment. Studies categorized as “semivaluable” (n=4, 17%) were flagged for issues such as the inappropriate use of qualitative methods to measure nonsubjective outcomes, suboptimal sample recruitment strategies, or insufficient consideration of bias. Although quality assessment was not an inclusion criterion for this systematic review, it was conducted to provide an overview of the quality of the eligible literature. Consequently, studies rated as “semivaluable” were still included in the data analysis. While these studies were limited in methodological rigor and offered less robust insights into the tools being evaluated, they contributed unique perspectives on health care workers’ trust in AI-CDSS tools that were not captured in other included studies. The included studies also discussed several limitations, including small sample sizes and various biases, such as potential selection bias, cognitive biases (eg, anchoring bias), and interviewer bias commonly associated with qualitative research. Other limitations included inaccuracies in self-reported data, regional differences in AI exposure, participants’ familiarity with the study context, and a focus on specific AI solutions or decision domains, all of which may limit the generalizability of findings.

Of the 27 included articles, summarized in Table 2 and Figure S1 in Multimedia Appendix 2, most were published recently: 12 (44%) in 2023, followed by 8 (30%) in 2022, 4 (15%) in 2024, and 3 (11%) in 2021. Geographically, most studies were conducted in the United States (n=12, 44%), followed by Europe (n=7, 26%), multinational collaborations (n=3, 11%), the United Kingdom (n=2, 7%), and 1 (4%) study each from China, Singapore, and Egypt. Most studies were conducted in hospital settings (n=17, 63%), across departments such as emergency care, radiology, and oncology, while 5 (19%) studies were conducted in primary care settings and 5 (19%) studies spanned both hospital and primary care environments. Study designs included qualitative research (n=16, 59%), mixed-methods studies (n=6, 22%), quantitative cross-sectional surveys (n=4, 15%), and 1 (4%) comparative evaluation study assessing AI-generated versus human-generated suggestions for clinical decision support. The study populations encompassed a wide range of health care providers such as physicians, nurses, nurse practitioners, general practitioners, intensive care unit clinicians, pharmacists, ophthalmologists, oncologists, interdisciplinary teams, behavioral health specialists, and AI practitioners. Sample sizes varied from small focus groups to cohorts exceeding 1000 individuals.

The included studies featured a wide range of AI-CDSS tools, demonstrating their application across various clinical functions and specialties (Table 2). These systems employ advanced technologies such as machine learning, deep learning, reinforcement learning, and explainable AI to support diagnostics, treatment planning, and clinical decision-making. Examples include the AI-based blood utilization calculator for improving transfusion procedures, the Brilliant Doctor system for dermatological diagnosis, machine learning and reinforcement learning models for providing sepsis treatment recommendations in intensive care unit settings, and QRhythm for identifying optimal rhythm management strategies in atrial fibrillation. Additional innovations are AI-CDSS tools for detecting and managing diabetic retinopathy, glaucoma, and cataracts; ChatGPT-enhanced electronic health record alerts for medication optimization in diabetes; as well as systems for lung cancer relapse prediction, vancomycin dosing, cardiovascular risk prediction, trauma radiography, and asthma management.

To analyze health care workers’ trust in AI-CDSS tools, we identified the methods used across 27 studies to assess trust-related elements. The majority employed semistructured qualitative interviews (n=19, 70%), followed by Likert scales (n=4, 15%), focus group discussions (n=1, 4%), How-Might-We questions (n=1, 4%), Likert scales combined with dichotomous questions (n=1, 4%), and the think-aloud protocol (n=1, 4%; Table 2). The assessment of trust in AI-CDSS tools revealed various factors that must be addressed to strengthen trust among health care workers (Table 2). Factors described in the study include experience with the AI system, colleagues’ recommendations, results from randomized controlled trials, and clinicians’ direct experience with the system over time. Transparency, accuracy, and the reliability of AI recommendations were identified as critical, with recurring concerns about the “black-box” nature of algorithms and the lack of clarity regarding how insights are generated. Additional factors influencing trust included perceived or actual risks, ease of use, organizational fit, and alignment with clinical judgment. Health care workers emphasized the importance of adequate training, customizable features, and the credibility of system developers. Trust perceptions were also shaped by considerations such as workload impact, acceptable error thresholds, and concerns around medical liability. Stakeholder involvement and familiarity with AI systems were described as contributing positively to trust in AI-CDSS tools. However, concerns about job displacement and the potential dehumanization of care emerged as significant challenges to fostering trust in these technologies.

The synthesis of study findings on trust in AI-CDSSs revealed 8 key thematic insights (Table 3 and Figure 2). These include (1) system transparency, which emphasizes the need for clear and interpretable AI systems; (2) training and familiarity, which highlights the importance of educating and familiarizing health care workers with AI-CDSS; (3) system usability, which focuses on seamless integration into clinical workflows; (4) clinical reliability, which stresses the need for consistent and accurate system performance; (5) credibility and validation, which describe the importance of system validation across diverse clinical contexts; (6) ethical consideration, which examines issues such as medicolegal liability, fairness, and adherence to ethical standards ,(7) human centric design, which focus on piroitizing patient centered approaches in design and finally, (8) customization and control, which reflect the need for AI tools to adapt to specific clinical needs while ensuring health care providers retain decision-making autonomy.

These themes were explored through the enablers and barriers that influence health care workers’ trust in AI-CDSS. Among the enablers, prior system use and validation through randomized controlled trials were cited as key factors that boosted confidence in the AI systems. Familiarity and training with AI tools further strengthened clinicians’ trust, empowering them to make informed decisions. Additionally, observing the system’s performance over time and receiving endorsements from colleagues and domain experts contributed significantly to trust building. Furthermore, system usability, alignment with clinical judgment, and the ability to reduce workload emerged as important factors positively influencing trust. Transparency in the AI development process and the perceived credibility of the developers also played a critical role in fostering confidence. Finally, the explainability and interpretability of AI recommendations, along with the ability to customize the system and seek clarification, offered clinicians a greater sense of control, further enhancing trust.

However, the study revealed several barriers that erode trust in AI-CDSS. A major concern was the black-box nature of the AI algorithm, which renders recommendations opaque and difficult to interpret. Clinicians also expressed concerns about inadequate training, which diminishes their understanding and confidence in using these systems effectively. Additional barriers included workflow disruptions, perceived threats to professional autonomy, and doubts regarding the accuracy and reliability of AI-generated recommendations. Ethical considerations, such as fears of dehumanization in patient care and perceived risks of job replacement, added further complexity to trust in AI-CDSS. Concerns were also raised about the efficacy of these systems, particularly due to inadequate external validation and limited generalizability to diverse clinical contexts. Lastly, trust was further undermined by unresolved issues related to medical liability, potential algorithmic biases, and broader ethical risks.

The systematic review included 27 studies analyzing health care workers’ trust in AI-CDSSs. The article selection process began with 333 records and, after rigorous screening based on inclusion criteria, was narrowed to 27 studies. Most studies were recent (n=12, 44% from 2023) and conducted in hospital settings across diverse health care worker groups. Qualitative methods dominated (n=16, 59%), with sample sizes ranging from small focus groups to cohorts of over 1000 participants. The synthesis of findings highlights 8 thematic areas influencing health care workers’ trust in AI-CDSS tools, encompassing both enablers and barriers. Key enablers include prior system validation, transparency, training, usability, and alignment with clinical judgment. By contrast, barriers such as algorithmic opacity, inadequate training, workflow disruptions, and ethical concerns undermine trust. Based on these themes, we provide actionable recommendations for the design and implementation of AI systems that are more likely to be trusted and accepted by health care practitioners (Table 3).

Fostering trust in AI-CDSSs among health care workers involves enhancing the transparency of AI algorithms and providing clear, practical, and actionable recommendations for clinical decision-making [39-41]. According to Nasarian et al [42], black-box models pose challenges in CDSS due to their limited interpretability, especially when compared with simpler white-box models that offer transparent results without requiring additional parameters. While black-box models often deliver high accuracy, their opacity can lead to confusion about how decisions are made—and, in some cases, may result in overreliance on the system, particularly among less experienced clinicians who may lack the expertise to interpret AI outputs effectively [22]. Gray-box models, positioned between the extremes of black-box and white-box models, offer a balance between complexity and interpretability, provided they are designed effectively [42]. Interpretability should be incorporated throughout the entire process, from data preprocessing and model selection to postmodeling phases. However, most existing AI-CDSS tools focus primarily on postmodeling explainability. To reduce skepticism surrounding the “black-box” nature of AI systems, developers should ensure transparency in the rationale behind recommendations at every stage of the development pipeline [42].

To improve trust in AI-CDSS, comprehensive training programs on AI tools for health care workers play a vital role [39,43]. These programs not only build familiarity with the systems but also enhance users’ confidence in their reliability and functionality. Dlugatch et al [6] discussed the impact of AI on health care workers. As AI technology begins to surpass human capabilities, the epistemic authority of medical practitioners risks being undermined, challenged, or even supplanted [6]. This may lead health care professionals to view AI-CDSS tools as replacements rather than assistants. To address these challenges, training programs should educate health care workers on how AI systems are developed, including their capabilities, limitations, and potential pitfalls.

To improve system usability and alignment with clinical judgment, hands-on training and peer-led workshops should be conducted [39,43]. These approaches not only enhance health care workers’ understanding of AI systems but also improve their practical usability. According to Task Technology-Fit (TTF) theory, users are more likely to adopt a technology only if it aligns with their tasks and improves performance. However, external influences and uncertainty surrounding AI can introduce biases that either encourage or deter clinicians from adopting such technologies in the future [25]. This highlight the importance of peer-to-peer sharing of experiences with AI-CDSS tools.

To ensure clinical reliability, AI systems should demonstrate accuracy and consistency through real-world testing and randomized controlled trials [44-46]. Micocci et al [22] noted that AI systems, like human clinicians, are inherently imperfect and should be designed to complement, not replace, the clinician’s holistic understanding of each clinical scenario. While AI can offer valuable decision support, the ultimate responsibility for diagnostic resilience lies with the clinician, who retains the authority to accept or reject AI recommendations [22].

Trust in AI-CDSS can be further fostered through external validation of the system in diverse clinical settings, which can help demonstrate the soundness of the AI methodology used in its development [47-49]. Nair et al [35] mentioned that clinicians express fatigue from the integration of AI-based tools into workflows, especially when organizations are reluctant to discontinue ineffective technologies. This underscores the crucial role of tool developers in thoughtfully managing and thoroughly validating systems across varied contexts to avoid adding further burden to health care workers.

The importance of human-centric design cannot be overstated in fostering trust in AI-CDSS [50,51,52,53,54, 55,56 ]. Amann et al [13] raised concerns about technology-induced dehumanization in patient care and its impact on the patient-clinician relationship. Sivaraman et al [28] and Jacobs et al [5] discussed the important role of a sociotechnical lens in designing AI-CDSS, emphasizing the need to integrate environmental and social factors into system development. Furthermore, Alruwaili et al [8] discussed that health care professionals, such as nurses, have varying concerns about AI’s impact on the human aspect of care, while others recognize its potential benefits. This highlights the importance of incorporating humanistic elements in the design of AI-CDSS as supportive tools that enhance, rather than detract from, patient care.

Clear guidelines on roles and responsibilities, along with robust validation of AI tools, will address liability related to ethical concerns, which will help alleviate concerns and build trust [57-61 ]. Gunasekeran et al [24] and Jones et al [5] noted that health care workers fear the medicolegal impact of AI-CDSS systems. Providing explicit guidance on the capabilities of AI-CDSS and clearly delineating the roles and responsibilities of health care workers can further help mitigate concerns related to medical errors and liability. Ethical AI frameworks, such as the European Commission’s Ethics Guidelines for Trustworthy AI, the EU AI Act, and the OECD’s AI Ethics Guidelines, offer specific guidance for the development of AI-CDSS in the health care sector [62-64]. These frameworks not only help reduce ethical concerns but also promote human-centric design, which can enhance health care workers’ trust in AI-CDSS.

Trust in AI-CDSS can be fostered through collaboration, coordination, and meaningful stakeholder engagement during system design, helping to eliminate ethical concerns and fears of job replacement among health care workers [65-72]. Chiang et al [12] emphasized the importance of securing support from a variety of stakeholders, such as organizational leadership and end users, early in the development process to improve trust in AI-based tools. Ball et al [73] also highlighted the role of collaboration and continuous communication through a “human-in-the-loop” approach, which integrates human expertise and addresses the limitations of AI algorithms. Involving direct end users, such as health care workers, during the development phase enables them to better understand the supportive role of AI-CDSS, rather than perceiving it as a threat to their jobs. Furthermore, engaging a range of stakeholders can help reduce ethical concerns by raising possible issues, such as harm to patients, early in the process. This, in turn, allows developers to make necessary modifications and improve trust in the implementation of AI-CDSS tools [11].

This systematic review has certain limitations. First, it included only studies published in English and did not account for AI system studies from nonindexed journals, which may limit the relevance of the findings to non–English-speaking or unpublished research. Second, the quality assessment was conducted using the CASP checklist, which evaluates only the qualitative elements of included studies, regardless of their overall design. This may have limited the generalizability of findings derived from nonqualitative studies. Third, although we aimed to conduct a meta-analysis, including a subanalysis or comparative analysis, we were unable to do so due to the high level of heterogeneity across studies and the lack of detailed demographic information. Additionally, sufficient data on health care roles and geographic contexts associated with the qualitative quotes and outcomes hindered our ability to conduct a comparative analysis. Lastly, the study was formative in nature, with categories and components generated through a subjective synthesis process, which may introduce interpretive bias. Despite these limitations, the synthesis and recommendations from this study help bridge existing gaps and provide specific themes to foster health care workers’ trust in AI-CDSS. Future studies should consider incorporating more diverse demographic data, performing cross-cultural studies, and exploring contextual differences in trust across various health care professional groups to address these gaps more comprehensively.

Our systematic review of 27 studies identifies 8 key themes influencing health care workers’ trust in AI-CDSS tools. We highlight important enabling factors such as transparency, training, usability, clinical reliability, and alignment with clinical judgment. Conversely, barriers include algorithmic obscurity, inadequate training, and ethical concerns. Based on these findings, we recommend prioritizing the development of transparent AI models, implementing comprehensive training initiatives, and conducting practical workshops with real-life testing to foster sustained trust in AI-CDSS among health care workers. Moreover, integrating human-centered design and addressing ethical considerations are crucial to ensuring that AI tools enhance, rather than hinder, the patient-health care worker relationship. Despite limitations, such as the exclusion of non-English studies, heterogeneity in study designs, and a lack of detailed data, the analysis is limited in conducting further exploration. Nevertheless, it bridges existing gaps and provides specific themes to foster the trust of health care workers in AI-CDSS. The identified thematic areas, along with our recommendations, establish a foundation for forthcoming research and development of AI-based tools to ensure that AI-CDSS are efficient, reliable, and trustworthy for health care workers.