Health care systems worldwide are increasingly strained, partly due to aging populations and insufficient resources, and there is increased demand for accessibility, medical quality, and economic efficiency [1]. Primary care is regarded as a cornerstone in health care systems across many regions of the world [2], and primary care clinicians’ job satisfaction is considered essential for many health care systems [3]. Recent studies have demonstrated the potential of artificial intelligence (AI) tools and systems to reduce burnout and increase the efficiency of health care professionals [4], as well as to improve diagnostic accuracy and patient care [5].

AI is an emerging technology with a broad range of applications [6-8]. However, there is still no consensus on a general definition of AI, which presents an obstacle to investigating peoples’ perspectives [9].

Recent advances in AI have led to increased health care–related AI use and research [10]. Previous reports have indicated that the main applications of AI in primary health care have been data extraction and processing [11], reducing administrative burden [12], and assisting physicians in diagnosing, determining a prognosis, and choosing a treatment [13]. Current large language models (LLMs) have started to play a more prominent role in health care, and new applications are frequently identified [14]. Several LLM products, including Chat Generative Pretrained Transformer (GPT), have demonstrated the capability of medical reasoning and have performed well on medical licensing exams [15,16]. Moreover, LLMs may improve communication between health care professionals and patients through text simplification [17].

Previous research suggests concerns among clinicians regarding the use of AI in health care, such as demographic biases, insufficient regulation, lack of trust in AI systems [18], and automation bias [19].

While there seems to be a lack of systematic synthesis on clinicians’ perspectives on AI in primary health care, a scoping review conducted in 2022 on perceptions and needs of AI in health care identified few studies within primary health care. End-user and stakeholder opinions are essential for future implementation and development. Since research on AI in primary care is limited and results are varied, perceptions of the use of AI in this domain are not fully understood [7].

The aim of this systematic review is to synthesize the current qualitative research on clinicians’ perspectives on AI in primary care settings.

We performed a systematic review and metasynthesis in accordance with the Cochrane Qualitative and Implementation Methods Group [26]. The review was reported according to the Enhanced Transparency in Reporting the Synthesis of Qualitative Research (ENTREQ) statement [27] (Checklist 1) and Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [28] (Checklist 2). The study protocol was registered with the International Prospective Register for Systematic Reviews (PROSPERO; CRD42024505209) [29] before conducting the review. This study employed a methodology for a systematic review of qualitative studies, in which the authors conducted a secondary qualitative synthesis of published clinician quotes and primary authors’ interpretations from the reviewed studies, allowing for deeper exploration of underlying patterns and themes.

Authors RB and AM imported the search results into Rayyan (citation manager) [32], where duplicates were removed. The authors independently screened titles and abstracts of the remaining articles against the inclusion and exclusion criteria. Any disagreements were discussed, and if consensus was not reached, a third author (AE) was consulted for a final decision. We included articles claiming to evaluate AI technology based on the authors’ definition of AI, as described in the “Introduction.”

Authors RB and AM independently conducted critical appraisal using the Critical Appraisal Skills Program (CASP) checklist for qualitative research [33]. Disagreements were discussed until a consensus was reached or author AE was consulted for a final decision.

Data were extracted from the Results section of the included articles and their supplementary material. Participant quotes and authors’ findings were analyzed independently by RB and AM to generate descriptive themes using thematic analysis according to the Braun and Clarke method [34]. This involved several steps through a primarily inductive analytic process. First, the authors familiarized themselves with the extracted data by reading it several times. RB then developed codes using line-by-line coding of words or sentences considered meaningful, using the NVivo software [35]. Data extraction and coding were performed in 2 stages. The first stage involved articles solely containing primary care clinician perspectives, and the second stage involved articles containing perspectives of both primary care clinicians and nonclinical health care professionals. Qualitative data with perspectives other than that of clinicians was not coded. Codes were discussed by both authors until an agreement was reached, whereafter, they were exported to a Microsoft Excel spreadsheet. RB then proceeded to generate descriptive themes by grouping codes. The alignment of codes to certain themes was discussed, and the descriptive themes were refined. Thematic synthesis, according to the Thomas and Harden method, was employed to develop higher-order analytical themes. It is a well-suited method for exploring qualitative data such as perspectives or sentiments [36]. Thematic synthesis was accomplished through a discussion between both authors, during which the analytical themes were developed and named (Multimedia Appendix 2). No new themes emerged from coding the articles with mixed perspectives. This method was chosen due to its ability to identify recurring themes and patterns across multiple studies, enhancing the breadth of the analysis.

Confidence in each synthesized finding was assessed using the Grading of Recommendations Assessment, Development and Evaluation-Confidence in Evidence from Reviews of Qualitative Research (GRADE-CERQual) approach. The GRADE-CERQual approach was chosen as it explicitly addresses qualitative evidence synthesis, allowing systematic and transparent assessments of the confidence in each thematic finding. Authors RB and AM independently evaluated each finding based on 4 components: methodological limitations, coherence, adequacy, and relevance. Each component was assessed as having no or very minor, minor, moderate, or serious concerns. Discrepancies were discussed, and if agreement was not reached, author AE was consulted for a final decision. Each finding began with an initial rating of “high confidence”. Confidence levels were then potentially downgraded to moderate, low, or very low based on the severity and number of concerns present in each component. Typically, one level of downgrading (eg, from high to moderate confidence) was applied when moderate concerns were identified in one component combined with minor concerns in other components, and two levels (eg, from high to low confidence) were applied when serious concerns or multiple moderate concerns were present. When concerns were minor or very minor, no downgrading was performed [37-42].

Because this systematic review used only pre-existing data, ethical approval was not required.

The final search generated 1492 results, and 415/1492 (27.8%) duplicates were excluded. The remaining 1077/1492 (72.2%) articles were screened by title and abstract, and 54/1077 (5%) articles were retrieved in full text and evaluated, of which 42/54 (77.8%) were excluded based on the exclusion criteria. Finally, 12/54 (22.2%) articles were included from the screening, and 1 additional article was identified from the reference lists of the previously included articles, resulting in a total of 13 studies [43-55] (Figure 2).

The 13 included studies were conducted in 6 different countries. Australia was the most frequent location with 4/13 (30.8%) studies [45-47,49], followed by Canada with 3/13 (23.1%) studies [51-53] and the United States with 3/13 (23.1%) studies [43,44,50]. Sweden [55], the Netherlands [54], and Germany [48] each contributed 1/13 (7.7%) study. Many studies used semistructured interviews for data collection (6/13, 46.2%) [43,45,48,51,53,55]. Mixed methods were used in 3/13 (23.1%) studies [44,47,50], focus groups in 2/13 (15.4%) studies [49,54], deliberative dialogue in 1/13 (7.7%) study [52], and a co-design workshop in 1/13 (7.7%) study [46]. Characteristics of the included studies are presented in Table 1.

The critical appraisal using the CASP qualitative checklist indicated that all included studies had clear research aims, appropriate qualitative methodologies, and well-reported findings. Several studies lacked sufficient transparency in ethical considerations (Multimedia Appendix 3).

The thematic analysis identified eight descriptive themes. These descriptive themes represent recurring elements identified across studies and served as the foundation for the synthesis of broader analytical themes. Thematic synthesis resulted in three analytical themes: the human-machine relationship, the technologically enhanced clinic, and the societal impact of AI. All themes are presented in Table 2. There was a wide time span across the included studies. Earlier studies, conducted between 2003 and 2013, explored perspectives on less advanced systems, such as CDSSs not based on ML [44,47,49], whereas later studies included more advanced AI systems, such as GAI and NLP [43,45,46,48,50-54], or a CDSS based on ML [55].

The results of the GRADE-CERQual assessment for the review findings are summarized in Table 3.

Details are provided in Multimedia Appendix 4.

When synthesizing primary care clinician perspectives of various AI systems, 3 analytical themes emerged. The GRADE-CERQual assessment indicated high confidence in 15 findings, moderate confidence in 5 findings, and low confidence in one finding.

There were many positive remarks on the potential for AI to assist clinicians in administrative tasks [45-47,51-53,55], clinical work [45,46,52,55], and education [51,53,54]. Some studies highlighted fears of AI replacing human roles [46,48,50]. Resistance to AI was noted, with concerns that AI might disrupt workflow and increase task time [43,47,48,50,53]. The underlying causes of skepticism were discussed. Some attributed it to previous negative experiences with other digital tools [51,52], others to their age or technical alignment [45,48,50]. Positive views on AI systems were generally seen in studies where AI tools were deemed effective, seamlessly integrated, and saved time [44,45,47].

Some clinicians saw benefits in automating tasks like documentation and consultation [44-48], whereas others feared potentially increased workload [43,47,50,53]. Technical concerns were found, including fears of computer errors or user errors [43-53,55]. Clinicians valued AI’s potential to assist in clinical tasks such as decision support [44,45,47-49,53-55]. Clinicians discussed desirable AI features, such as diagnostic support, integration with telemedicine, and customization options [44-52,55]. Several studies mentioned the importance of the system’s user-friendliness [47-49,51,55].

Security and privacy issues were highlighted, particularly regarding the handling of sensitive patient data and the risks of unauthorized access [43,45-48,50]. Societal impact, including potential biases and overreliance on AI, was concerns [45,46,48,50-52], and potential legal implications if clinicians acted outside of the AI recommendations [45,46,50,55]. Clinicians expressed greater confidence in adopting AI systems that had received formal regulatory approval or institutional endorsement [49,51].

Our findings regarding clinicians’ views of the potential of AI and reservations regarding safety aspects are similar to a previous systematic review of AI-powered chatbots for managing chronic illness, which provided insights into the usability and acceptance of AI in health care. The review found that participants gave positive feedback regarding perceived usefulness, satisfaction, and ease of use. The review also concluded that the safety of AI-powered chatbots has been overlooked and needs to be considered more thoroughly in future designs [56].

As AI expands into health care, a significant concern has emerged: the risk of bias. Since AI relies on historical data that could be statistically or socially biased, it could potentially incur a risk of worsening patient outcomes [57]. This coincides with our findings regarding clinicians’ concerns about biased AI systems.

In this review, clinicians were positive toward simplifying certain tasks using AI while simultaneously having concerns regarding technical aspects. Another systematic review of stakeholders’ perspectives on clinical AI implementation, which included perspectives of health care providers, similarly found that health care providers saw benefits in using AI for reducing repetitive tasks, improving patient outcomes, and clinical training. Reservations toward AI included implementation issues, uncertainty around its mechanics, and skepticism toward its ability to inform clinical decisions [58].

In the current review, opinions on whether clinicians should be involved in the development of AI systems were mixed. One scoping review published in 2020 found the engagement of clinicians in health care AI development and research to be crucial [59]. The ethical and regulatory challenges expressed in the findings of this review are also brought to light in a scoping review published in 2022. Their findings suggest that AI research and development in health care is currently outpacing the creation of supporting AI governance, and there is a need for international collaboration to facilitate comprehensive AI governance in this sector [60]. There were similar findings in another article published in 2021, where the author concludes that there is an apparent risk of regulations and oversight falling behind AI’s rapid development and integration [61].

This review focuses solely on clinicians, although many other professionals are implicated in the adoption of AI into health care. Further investigation of perspectives of information technology experts, managers, and other stakeholders could prove valuable in the development, adoption, and integration of AI systems [62].

The findings of this review suggest that AI systems should focus on reducing administrative burden and supporting certain clinical tasks, provided they do not disrupt clinicians’ workflows. These systems must demonstrate time-saving capabilities and seamlessly integrate with existing infrastructure, such as electronic medical records. Through these functions, an AI system could enhance a medical visit by allowing the clinician to focus on patient contact rather than administrative tasks.

Continuous monitoring for computer errors, structured AI training programs for clinicians, and simplifying user interfaces are essential to minimize user errors. Additionally, ensuring robust data handling practices is critical to maintaining patient privacy and security. There is also a desire among clinicians for clearly defined medicolegal responsibilities.

Developers of health care-related AI systems should aim to mitigate system bias and consider collaborating with clinicians in the development process to build initial trust and address potential concerns. Involving clinicians with previous experiences of AI or CDSS in the development or integration of AI systems might facilitate adoption and use. Moreover, current AI tools should complement, not replace, clinical decision-making. It is important to provide younger clinicians with opportunities to develop critical reasoning skills without fostering over-reliance on AI-generated outputs.

The review revealed gaps regarding clinician perspectives on AI in primary health care, specifically in LLMs. Perspectives on ethical implications focused mainly on bias in AI systems, patient privacy and data security, medicolegal implications, transparency and accountability, and equity in AI system access. However, deeper analyses regarding the ethical implications of modern AI systems, including how AI might alter clinicians’ professional roles, authority, and autonomy, were scarce. Some of the included studies mentioned clinicians expressing thoughts on their reliance on AI systems and how their autonomy might be affected. However, further research could provide valuable information on these aspects that directly influence clinicians’ acceptance and utilization of AI technologies.

Findings from this review suggest that some clinicians are aware of long-term job implications and possible job displacement due to the introduction of AI into health care. Threats to the professional autonomy of clinicians could be due to automation bias, potentially overriding or deskilling clinical judgment with decision-support recommendations becoming default options, or by reducing the professional freedom of clinicians when many clinical actions are turned into data, enabling scrutiny of even the smallest decisions. Further research on AI’s effects on the evolution of clinicians’ career paths and future autonomy is warranted.

While many clinicians assume ultimate responsibility in patient care, several fear “legal whiplash” if they disregard an AI recommendation that later proved correct or followed one that proved harmful. These findings highlight the need for clear governance frameworks by having AI tools accredited and liability boundaries specified by a professional body such as the EU AI Act [63] and FDA Software-as-a-Medical-Device (SaMD) guidance [64]. Further empirical research is recommended to evaluate how such regulations translate into everyday primary care. Another possible pre-condition for ethically acceptable AI use could be provided by transparent AI reasoning paths, enabling clinicians to follow a defensible audit trail rather than a simple, final output that could prevent shared decision-making with patients.

Further research on AI in primary health care is needed, especially in low- and middle-income countries. Notable gaps in the literature include evaluations of LLMs in primary health care, which are expected to have great potential. Ethnographic studies could yield deeper insights into AI’s impact on the professional role of clinicians and long-term career implications. Further in-depth, qualitative research on these topics could prove helpful for future AI system development and integration.

Future research should also expand beyond cross-sectional studies to longitudinal, mixed-methods studies that follow AI systems from adoption to routine use in primary care clinics for further in-depth analysis of AI use and a deeper understanding of facilitators and barriers to adoption. Additionally, systematic reviews targeting specific types of AI or clinical use cases would support a more nuanced understanding of AI implementation in diverse primary care contexts.

Clinicians view AI as a technology that could both enhance and complicate primary health care. While AI can provide substantial support, its integration into health care requires careful consideration of ethical implications, technical reliability, and the maintenance of human oversight. Interpretation is constrained by heterogeneity in qualitative methods and the diverse AI technologies examined across studies. More in-depth qualitative research on the effects of AI on clinicians’ careers and autonomy could prove helpful for the future development of AI systems.