AI is rapidly emerging as a transformative force in modern medicine and is promising for reshaping clinical workflows, supporting personalized care, and improving patient outcomes [1,2]. Although its application has historically focused on image-centric fields such as radiology and dermatology [3,4], AI is increasingly permeating the management of chronic conditions that require continuous data analysis and complex decision-making processes [5]. Among these, diabetes mellitus (DM) is particularly well suited to AI-supported care because its daily management requires the continuous interpretation of large volumes of data, including glycemic patterns, insulin dosing, dietary intake, physical activity, and other behavioral and physiological variables [6-8].

In recent years, diabetes care has evolved from reactive monitoring to proactive and increasingly automated management approaches [9]. AI-driven technologies, particularly machine learning (ML) algorithms, now support advanced tools such as hybrid closed-loop systems, which automate insulin delivery in response to real-time glucose sensor data [10,11]. AI models are also being applied beyond therapeutic delivery to risk prediction. This offers the potential to predict adverse events such as hypoglycemia or diabetic ketoacidosis before they occur and to identify individuals at high risk of disease progression [12]. Collectively, these advancements may substantially reduce the cognitive and emotional burden of diabetes self-management, which is often described as a full-time job for patients and their caregivers [13].

Despite the enthusiasm, the transition from theoretical models to clinical practice has revealed important limitations, as algorithmic precision does not necessarily ensure clinical effectiveness or safety. Evidence indicates that automated systems are susceptible to routine human errors and raise complex ethical concerns regarding data privacy and the medicalization of domestic life [14,15]. Furthermore, the successful implementation of these innovations requires a clear understanding of 3 key dimensions: their benefits for individuals, the barriers that limit their adoption, and their potential future applications [16]. Currently, many ML models have not demonstrated clear clinical superiority over traditional statistical methods and remain constrained by a lack of external validation and poor generalizability [6].

The successful adoption of AI in diabetes management depends not only on algorithmic performance but also on the trust and acceptance of the stakeholders involved. Although numerous quantitative studies have evaluated the clinical efficacy of these tools [17,18], evidence regarding human experience, specifically the perceptions, barriers, and facilitators of their adoption, remains fragmented and often limited to specific devices or user groups [19].

The primary objective of this systematic review was to synthesize qualitative evidence to provide an in-depth understanding of the perspectives of patients, caregivers, and health care professionals (HCPs) regarding the use of AI in diabetes management, glycemic control, and risk prevention.

This systematic review was performed in accordance with the methodological framework established by the Cochrane Qualitative and Implementation Methods Group (CAMELOT) [20], and the protocol was prospectively registered in the International Prospective Register of Systematic Reviews (PROSPERO) database (registration number: CRD420251010095). The reporting of this review complies with the Enhancing Transparency in Reporting the Synthesis of Qualitative Research (ENTREQ) statement (Multimedia Appendix 1) [21].

A systematic literature search was conducted by 2 reviewers (HM-M and JM-A) across 5 major databases: MEDLINE (via PubMed), Web of Science, Scopus, CINAHL, and PsycInfo, covering the period from their inception to February 17, 2026. To develop the search strategy, we used the Sample, Phenomenon of Interest, Design, Evaluation, and Research type (SPIDER) framework (Multimedia Appendix 2) [22] to identify relevant qualitative and mixed methods research.

The search terms combined concepts related to diabetes management, AI, and user perceptions. The specific search syntax was tailored to the requirements of each database (Multimedia Appendix 3). Eligible studies were required to meet the following criteria: (i) provide a qualitative analysis of perceptions, barriers, or facilitators related to the use of AI tools for diabetes management or risk of complications; (ii) include adult patients, caregivers, or HCPs as the study population; and (iii) use qualitative data collection methods, such as semistructured interviews or focus groups, requiring a formal qualitative data analysis approach. For mixed methods studies, inclusion depended on the availability of separate, distinct qualitative findings. Only articles published in English or Spanish were considered.

For this review, an AI-enabled tool was operationally defined as a digital system that the primary study explicitly described as using AI, ML, deep learning, natural language processing, generative AI, or an adaptive data-driven algorithm to perform prediction, classification, recommendation, personalization, or decision support beyond fixed-rule data capture, storage, display, or communication. Apps and wearable devices were eligible only when an identifiable AI-enabled function was integral to the phenomenon explored; general mobile health (mHealth), telehealth, or monitoring tools without an AI component were excluded. Studies that focused exclusively on the diagnosis of complications through images, such as diabetic retinopathy, were excluded because this review focused on AI embedded in ongoing diabetes management, self-management, and clinical decision support. Image interpretation represents a distinct diagnostic pathway with different users and workflows that warrants separate synthesis. However, studies that included image analysis as one component of a broader diabetes-management tool remained eligible. Studies focusing solely on technical validation, as well as gray literature, conference abstracts, protocols, dissertations, and theses, were excluded.

Following the search, the citations were imported into Rayyan software (Rayyan Systems Inc) [23], and duplicates were removed. The screening process was conducted in 2 stages by 2 independent reviewers (HM-M and JM-A). First, titles and abstracts were assessed against the inclusion criteria. The full texts of potentially relevant records were subsequently assessed to determine final eligibility. Any discrepancies arising during this process were resolved through discussion or, if necessary, by consulting a third reviewer (VM-V).

The methodological rigor and risk of bias of the included studies were evaluated independently by 2 reviewers (HM-M and JM-A), with disagreements resolved through consensus. To ensure a comprehensive assessment, 2 complementary tools were used. First, the Joanna Briggs Institute (JBI) critical appraisal checklist for qualitative research (Table 1) was used to assess the trustworthiness of the studies [24]. Additionally, the CAMELOT tool (Multimedia Appendix 4) [25] was applied to identify specific limitations, which informed the subsequent Grading of Recommendations Assessment, Development, and Evaluation–Confidence in the Evidence From Reviews of Qualitative Research (GRADE-CERQual) assessment of confidence in the review findings [26].

Two reviewers (HM-M and BS-R-H) independently extracted data using a standardized form. Extracted variables included study characteristics, participant characteristics, diabetes type, stakeholder role, data collection and analysis methods, AI function, level of participant exposure, and the analytical themes to which each study contributed (Table 2 and Multimedia Appendix 5). Raw participant numbers were not used to weight study contributions. Instead, contribution to the synthesis was assessed according to relevance, contextual richness, and depth of qualitative data. Open-ended questionnaire responses were therefore treated as less contextually rich than interview, focus-group, or longitudinal co-design data when assessing adequacy.

We adopted a thematic synthesis approach [41] to integrate the qualitative findings. Two reviewers (HM-M and JM-A) performed line-by-line coding of the findings reported in the primary studies. These codes were initially organized into descriptive themes grounded in the data. These themes were subsequently interpreted to generate overarching analytical themes that provided new insights into the phenomenon under study. Disagreements were resolved through discussion and, when necessary, arbitration by a third reviewer (VM-V). To examine the influence of questionnaire-based qualitative evidence, we repeated the study-to-theme mapping after excluding the cross-sectional study in which qualitative data consisted of open-ended questionnaire responses [39].

Confidence in each synthesized finding was appraised using GRADE-CERQual (Table 3 and Multimedia Appendix 6) [26,42].

Two reviewers (HM-M and JM-A) independently assessed methodological limitations [43], coherence [44], data adequacy [45], and relevance [46] for the body of evidence contributing to each finding, with disagreements resolved by consensus and, when necessary, consultation with a third author (VM-V). The JBI and CAMELOT assessments were used to inform, rather than mechanically determine, the methodological limitations component of GRADE-CERQual. Concerns were considered at the level of each synthesized finding according to which studies contributed data to that finding. The absence of reported researcher positioning or reflexivity was treated as a recurrent limitation across the evidence base. Ratings reflected the richness, quantity, and distribution of contributing data rather than sample size alone. Potential concerns related to sampling, demographic imbalance, transferability, or exposure to hypothetical or prototype technologies were considered within the methodological limitations, adequacy, or relevance components, as appropriate.

After removing duplicates, 602 records underwent title and abstract screening. Of these, 114 articles were selected for full-text review, and 14 [27-40] met the inclusion criteria (see Figure 1). These articles corresponded to 14 unique study populations.

A total of 14 studies published between 2023 and 2025 were included in this review (Table 2) [27-40]. Three included studies were conducted in Singapore [27,31,34], 2 in the United States [28,32], 2 in Saudi Arabia [30,39], 2 in Malaysia [33,35], and 1 each in the United Arab Emirates [29], Belgium [36], Germany [37], the United Kingdom [38], and India [40]. Of these, 13 explicitly reported the number of participants [27-39], while one study did not specify the total number of interviewees [40]. Among the studies with available data, the total number of participants was 738. This sample comprised 598 patients with diabetes [27,30-32,35,37-39] and 140 other stakeholders, including health care professionals, hospital staff, and family members [27-29,32-34,36,40]. Participants’ ages ranged from 18 to 81 years, and the overall proportion of female participants was 41.9%.

With respect to data collection, most studies used semistructured interviews [27-29,31,33,35,37,40], focus groups [34], or a combination of interviews and focus groups [32]. Other designs included longitudinal co-design [38], a one-week direct-use study followed by interviews [30], an interactive prototype user study with open-ended feedback [36], and a structured questionnaire containing open-ended responses [39]. Study [39] contributed 418 of the 738 participants with known sample sizes; however, its large numerical sample was not interpreted as providing proportionally greater qualitative weight because its open-ended responses offered less contextual depth than interviews, focus groups, and longitudinal co-design.

The included literature was heterogeneous across stakeholder groups, diabetes types, AI functions, care settings, and levels of exposure (Multimedia Appendix 5). Direct-use studies tended to provide practical accounts of workflow fit, self-management support, and interaction burden, whereas prototype, wireframe, and hypothetical-system studies more often elicited expectations concerning desired features, explainability, trust, and anticipated risks. These patterns were interpreted descriptively because the number and design of studies within each stratum did not support formal comparative subgroup analysis. AI was therefore not treated as a single homogeneous intervention.

Methodological quality was assessed using CAMELOT and JBI Critical Appraisal Checklist for Qualitative Research.

A thematic synthesis of the 14 included studies [27-40] revealed 4 main themes and 13 subthemes related to stakeholder perspectives on the integration of AI in diabetes care (Figure 2). All relevant quotations are provided in the supporting information (Multimedia Appendix 7), organized systematically by themes and their corresponding subthemes.

Contributions differed across stakeholder and diabetes groups. Patients most frequently contributed to findings concerning self-management, emotional and practical burden, usability, and access to human support. Health care professionals more often emphasized preventive utility, workflow integration, accuracy, professional autonomy, accountability, and explainability. Family caregivers primarily contributed to shared monitoring, communication, and caregiver burden. Type 1 diabetes mellitus (T1DM)–focused studies particularly addressed glucose variability, insulin decisions, unexpected situations, and device integration, whereas type 2 diabetes mellitus (T2DM)–focused studies more often addressed behavioral support, education, family involvement, risk prediction, and prescribing. The study-level distribution of these contributions is presented in Multimedia Appendix 5.

The sensitivity analysis excluding the questionnaire-based study [39] did not remove or materially alter any of the 4 themes or 13 subthemes. That study contributed supporting data to decision-making support and real-life adaptation, trust and data security, and usability and technical preferences, but none of these findings depended primarily on its data. The GRADE-CERQual confidence ratings were unchanged after its exclusion.

This systematic review of 14 qualitative studies [27-40] revealed ambivalent but predominantly positive stakeholder perceptions of AI in diabetes care. Supported largely by high-confidence GRADE-CERQual assessments, the findings suggest that patients and HCPs view AI not only as a diagnostic aid, but also as a potentially useful educational and motivational tool. A frequently perceived benefit was its capacity to support a shift from reactive treatment to preventive care [28,33,40]. Notably, this broader perspective was considered especially relevant for older adults and people living in underserved areas, where AI may help bridge critical gaps in health care access [47].

However, participants perceived adoption as limited by a persistent lack of trust, concerns over data privacy, and the potential for increased systemic and caregiver burden. Specifically, the continuous nature of diabetes monitoring led participants to anticipate alert fatigue, data logging burnout, and heightened anxiety when recommendations were overwhelming or lacked clinical accompaniment [27,28,31,33]. Fears that AI could undermine professional autonomy and the essential human touch in care were also prominent [29,34,36]. Concerns regarding algorithmic accuracy are consistent with findings from other domains of digital health, where the “black box” nature of AI triggers institutional skepticism [48]. Nevertheless, this review reveals nuances specific to diabetes care. Participants valued AI as a potential safety net during out-of-routine situations, where usual self-management strategies may fail [38]. While AI applications in other chronic pathologies often emphasize medication adherence [49], participants described diabetes care as requiring lifestyle personalization and management of a substantial mental load [30,31].

Unlike previous quantitative studies that focused mainly on clinicians’ perspectives [50], our findings incorporate the views of patients and caregivers. One notable contribution is the perception of AI as a family communication bridge [27], which underscores the importance of the social and relational functions of digital health tools [51]. This shifts the focus beyond individual monitoring toward supporting the wider network involved in diabetes care. Recent perspectives highlight that large language models (LLMs) offer significant potential for this type of personalized health coaching and nutritional guidance, yet they require rigorous validation and specialized fine-tuning to mitigate critical risks such as algorithmic hallucinations [52].

Stakeholders’ concerns about reliability, privacy, transparency, and cross-institutional deployment also have technical foundations. Multicenter modeling research illustrates efforts to address small samples, distribution differences between sites, knowledge calibration, privacy preservation, generalization, and interpretable classification [53,54]. These studies offer technical context for why participants’ concerns about representative data, explainability, and reliable performance across clinical settings are relevant to implementation.

The primary qualitative evidence synthesized in this review supports that stakeholders prioritize specific technical and systemic requirements for AI adoption. Findings derived directly from the included studies emphasize that usability must prioritize seamless integration with continuous monitoring systems and wearable devices to minimize the burden of manual data entry and mitigate logging fatigue [37,39,40]. Furthermore, stakeholders explicitly demanded transparency, explainability mechanisms, and tailored, multichannel education models to support equitable implementation, particularly for older adults and individuals with lower digital literacy [28,32,33,36].

Broader requirements concerning regulatory classification, formal liability allocation, and institutional governance were not directly evaluated in most of the qualitative studies included. Our interpretation of these issues is informed by external policy frameworks, specifically the World Health Organization (WHO) guidance on AI for health and the European Union Artificial Intelligence Act (EU AI Act) [55,56]. These documents provide the necessary regulatory context regarding transparency, intelligibility, accountability, autonomy, and the protection of sensitive data for high-risk health care systems.

A key strength of this review is its multistakeholder approach, which involves the synthesis of perspectives not only from HCPs but also from patients, caregivers, and administrators. By integrating these diverse viewpoints, this review provides a more comprehensive understanding of the sociotechnical factors that shape the implementation of AI in diabetes care. Furthermore, the use of the JBI, CAMELOT, and GRADE-CERQual frameworks enhances the methodological rigor of the synthesis, enabling qualitative findings to be interpreted with high transparency and to inform implementation considerations, clinical practice discussions, or institutional policy development.

However, several limitations must be acknowledged. First, the methodological quality of the primary literature presented specific constraints. In particular, the JBI assessment revealed limited reporting of researcher reflexivity in several studies. This restricts the ability to fully evaluate how researcher positioning, professional backgrounds, and prior assumptions may have influenced data collection and thematic interpretation. Second, potential selection bias was identified in the underlying literature, particularly concerning gender imbalances [31,35,36,38] and overrepresentation of highly educated participants [39]. These demographic patterns may have influenced the findings by overrepresenting individuals who are more familiar with, confident in, or receptive to digital health technologies. Consequently, this review may reflect a more favorable perception of AI acceptability than would be found among populations with lower digital literacy, limited access to technology, or greater socioeconomic vulnerability. Third, the inclusion of only English and Spanish language publications may have excluded relevant studies from other linguistic, cultural, and health care contexts. This limitation could be particularly relevant because attitudes toward AI, data sharing, clinical authority, and self-management technologies are likely to vary across health systems and sociocultural settings. Fourth, it is important to note the nature of the synthesized evidence. These qualitative findings describe stakeholder experiences, expectations, and beliefs; they do not demonstrate clinical effectiveness, safety, diagnostic accuracy, or improved patient outcomes, which must be established through quantitative and experimental designs. Fifth, the evidence base was heterogeneous across stakeholder groups, diabetes types, AI modalities, care settings, and level of exposure. Some participants described direct experience with functioning tools, while others evaluated prototypes, wireframes, general concepts, or hypothetical systems. Perceptions based on anticipated use may not predict responses after sustained real-world exposure, and the synthesis should therefore not be read as a uniform estimate of acceptability across technologies or populations. Finally, because conversational LLMs and associated technologies are evolving rapidly, stakeholder perceptions may shift faster than conventional scientific publication timelines can capture.

Stakeholders perceived AI as offering possibilities for continuous, personalized support beyond traditional clinical encounters, as well as assisting with decision-making, patient empowerment, and more responsive models of care. However, these findings reflect qualitative perceptions, experiences, and expectations and should not be interpreted as evidence that AI improves clinical effectiveness, safety, or patient outcomes. Successful implementation will depend on addressing concerns regarding trust, transparency, accountability, equity, and the preservation of human-centered care. The future of AI in diabetes care should therefore be guided by transparent design, meaningful stakeholder involvement, and careful integration into existing clinical, social, and family support networks. Rather than replacing human relationships in care, AI should function as a supportive tool that strengthens communication, shared decision-making, and individualized diabetes management.