This scoping review was conducted in accordance with the recommendations of the Joanna Briggs Institute (JBI) [20], and was reported in accordance with the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) guidelines (Checklist 1) [19]. A protocol for this scoping review was prospectively registered with the Open Science Framework [21].

The selection criteria for this review were developed using the population, concept, and context framework. The population of interest was comprised of clinicians or health professionals working in mental health care, including psychiatrists, psychologists, mental health nurses, medical professionals, counselors, social workers, and other allied health staff. Studies containing mixed participant groups (eg, those reporting attitudes of clinicians and patients) were eligible when findings relating to clinicians or health professionals were reported separately and could be extracted.

The concept of interest was clinicians’ attitudes, perceptions, beliefs, knowledge, acceptability, readiness, concerns, or experiences of AI. Studies in which AI was examined broadly were eligible, including generative AI, machine learning, clinical decision support systems, conversational agents, documentation tools, screening or monitoring systems, and predictive applications. The context was mental health care, including psychiatric, psychological, counseling, psychotherapy, and other mental health service settings.

Studies were eligible for inclusion if they reported empirical data relating to clinicians’ attitudes, perceptions, or beliefs regarding the use of AI in mental health care. This included studies using surveys, interviews, focus groups, mixed methods designs, or other approaches that captured clinicians’ perspectives. Studies were included where AI was examined broadly, or where a specific type of AI or use case was examined.

Studies were excluded if they evaluated technical performance, diagnostic accuracy, or clinical effectiveness of an AI tool without reporting on clinicians’ attitudes, perceptions, or beliefs. Studies were also excluded if they focused exclusively on patients, students, caregivers, developers, or the general public, unless clinicians’ perspectives were separately reported. Reviews, conference abstracts and proceedings, errata, editorials, poster presentations, commentaries, non–peer-reviewed documents, and articles not available in English were also excluded. Studies published before January 2020 were excluded to capture the most recent evidence, reflecting the rapid development and increasing accessibility of AI technologies, particularly generative AI. Non–peer-reviewed sources were excluded as the review aimed to synthesize primary empirical evidence concerning clinicians’ perspectives. Studies not available in English were excluded, as resources for translation were not available. No restrictions were applied according to geographical location, clinician discipline, participant age group, or type of mental health service.

The final search was conducted on May 5, 2026, and included 6 databases searched individually: CINAHL (EBSCO), Embase (Elsevier), PsycINFO (Ovid), PubMed (National Library of Medicine), Scopus (Elsevier), and Web of Science (Clarivate), using both MeSH terms or other database-specific subject headings, and free-text keywords. The search was limited from January 1, 2020, to the date of the search using database filters, in order to capture the latest AI-focused research in this rapidly developing area. No additional registries, websites, print sources, or search methods were used.

The search strategy was developed by the authors, was not based on any previously published searches, and was reported in line with the PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Literature Search Extension) guidelines (Checklist 2). The search strategy was developed considering each aspect of the population-concept-context, in consultation with a research librarian, and was piloted and subsequently refined by the authors. The search encompassed key terms relating to (1) clinicians; (2) AI; (3) attitudes, perceptions, or beliefs; and (4) mental health.

The preliminary search strategy was developed in PubMed and piloted to assess the relevance of retrieved records. It was then translated for the remaining databases, with database-specific subject headings and syntax adapted to the requirements of each platform. Searches were executed by 1 author (CH). The full reproducible search strategies for all databases searched are available in Multimedia Appendix 1.

The results of the database searches were imported into EndNote (Clarivate), and then uploaded into Covidence, a systematic review management tool [22]. Articles were deduplicated by Covidence (Veritas Health Innovation), and then screened by 2 authors (CH and MR), first by title and abstract, then by full text.

Eligibility criteria were operationalized using a standardized screening form in Covidence available to both reviewers. Before formal screening, both reviewers independently assessed a pilot sample of titles and abstracts. Differences in the interpretation of the eligibility criteria were discussed, and the screening guidance was refined before the remaining records were assessed. Both reviewers independently screened each title and abstract. Records considered potentially eligible by either reviewer proceeded to full-text assessment. The reviewers then independently assessed each full-text article against the eligibility criteria and recorded a primary reason for exclusion for articles that were excluded. Conflicts were discussed between the raters until a consensus was reached. Interrater agreement was assessed using Cohen κ.

A data extraction template was developed in Microsoft Excel by 1 author (CH), which was piloted and refined in consultation with a second author (MR). The form was piloted using a subset of included studies selected to represent different study designs and AI applications. Following the pilot, the reviewers discussed the relevance of each field, and the form was refined to ensure consistency and that all data relevant to the review questions were captured.

Data items were extracted by 2 authors (CH and MR), who cross-checked results for accuracy and consistency. Discrepancies were resolved through discussion and consensus, with reference to the original publication. The charting form was updated iteratively where new relevant concepts emerged, and previously charted studies were reviewed following any changes to ensure consistent application across the dataset. Where information was not reported or could not be determined from the publication, it was recorded as “not reported.”

Extracted data included authors, year of publication, country, study designs, data collection methods, participants’ occupation, type of AI examined, study domains or topics explored, and findings relevant to clinicians’ attitudes, perceptions, and beliefs. Data were also extracted, where reported, on perceived benefits, risks, use cases, implementation considerations, governance issues, needs for education, impacts on professional roles, impacts on patients, and overall sentiments toward AI. Professional groups were initially recorded using the terminology reported by each study and were subsequently grouped into broader categories. AI applications were classified according to their described function. In both cases, categories were not mutually exclusive, and studies could contribute to more than one professional or AI application category.

Study characteristics were summarized descriptively using counts and percentages for publication year, geographical location, study design, participant profession, and type of AI application examined. Findings relating to clinicians’ attitudes, perceptions, and beliefs were synthesized narratively. Extracted findings were initially organized according to the review objectives and recurring concepts identified across the included studies. Similar findings were compared and grouped into broader descriptive domains through discussion between 2 reviewers (CH and MR). The developing domain structure was applied across all included studies and refined iteratively until it represented the range of perspectives reported.

The studies’ key findings were synthesized into nine domains: (1) AI knowledge, familiarity, and prior use; (2) attitudes, perceptions, and emotional responses; (3) perceived benefits and opportunities; (4) perceived risks, harms, and ethical concerns; (5) potential uses for AI in mental health care; (6) implementation, integration, and adoption readiness; (7) impact of AI on professional roles and future practice; (8) impact on the therapeutic relationship; and (9) education, training, policy, and governance needs.

The database searches returned a total of 12,356 records. Following the removal of 3467 duplicates, 8889 sources were screened by title and abstract, with 8786 being deemed irrelevant. The remaining 103 articles were sought for retrieval, with 15 of these being excluded due to the full text being unavailable. The remaining 88 full texts were screened, with 53 excluded as they did not meet the inclusion criteria for the following reasons: wrong population (n=24), wrong outcomes (n=13), wrong setting (n=9), wrong study design or publication type (n=3), not being AI-focused (n=2), and studies not available in English (n=2). The raters demonstrated moderate agreement when screening (Cohen κ=0.60). A total of 35 articles were included in the final review (Figure 1) [3,11,12,23-54]

The included studies were published between 2020 and 2026, with most (n=29, 82.9%) being published from 2024 onwards. Studies were conducted globally, targeting populations in North America (n=9, 25.7%) [30,35,36,41,46-48,51,54], the Middle East (n=7, 20.0%) [11,25-27,39,40,52], Europe (n=6, 17.1%) [28,34,44,45,49,53], Asia (n=4, 11.4%) [23,32,43,50], Australia (n=3, 8.6%) [3,12,37], and Africa (n=1, 2.9%) [24]. Four (11.4%) studies were conducted globally [29,33,38,42] and 1 study (2.9%) was conducted across both Germany and the United States [31] (Multimedia Appendix 2, Figure 2).

Eighteen (51.4%) of the studies used surveys [11,12,23-27,29-31,33-35,38,42,47,49,53], while 16 (45.7%) used either individual or focus group interviews [3,28,32,36,37,39-41,43-46,50-52,54]. One (2.9%) study used both a survey and an interview [48]. Participants were all individuals who worked clinically within a mental health care setting and included psychiatrists (n=20, 57.1%) [3,11,12,24-27,31,33,34,37-39,42-44,50,51], psychologists (n=14, 40.0%) [3,11,12,23,27,28,31,34,37,39,42,44,45,50], counselors or therapists (n=14, 40.0%) [3,12,31,34,35,37,39,41,42,45-47,49,53], nurses (n=9, 25.7%) [11,12,28,32,40,43,44,52,54], social workers or welfare officers (n=8, 22.9%) [3,12,28,37,42,46,50,54], medical doctors and general practitioners (n=7, 20.0%) [11,12,23,26,36,48,51], support workers (n=4, 11.4%) [3,12,37,42], allied health staff (n=2, 5.7%) [12,23], and other mental health professionals (unspecified; n=2, 5.7%) [11,42] (Multimedia Appendix 2).

The majority of studies discussed AI broadly (n=14, 40.0%) [11,12,23,25,26,29,32,33,35,42,51-54] or in terms of generative AI (GenAI) or large language models (LLMs) such as ChatGPT [55], Gemini [56], or Copilot [57] (n=14, 40.0%) [3,12,27,30,37-44,46,49]. Studies also discussed clinical decision support tools (eg, AI-supported diagnostic, treatment planning, risk detection, or triaging tools; n=12, 34.3%) [24,26,28,32,34,36,40,44,48,50-52], psychotherapy chatbots or conversational agents (n=9, 25.7%) [3,12,24,30,31,37,39,45,49], documentation or notetaking tools (n=4, 11.4%) [31,37,41,46], tools providing supervision or feedback on clinicians’ performance (n=4, 11.4%) [24,31,46,47], or symptom monitoring or prediction tools (n=3, 8.6%) [32,34,44] (Multimedia Appendix 2).

This scoping review aimed to synthesize the current evidence on clinicians’ attitudes, perceptions, and beliefs around the use of AI in mental health care. The findings indicate that clinicians are cautiously optimistic about AI, but their acceptance is conditional on how AI is used. Clinicians recognized potential value in AI tools, which could reduce administrative burden, support documentation, assist with screening or monitoring, and improve access to care, but raised concerns about privacy, clinical safety, accountability, professional role boundaries, and impacts on the therapeutic relationship. Overall, clinicians do not necessarily reject AI in mental health care, but successful implementation requires clear role boundaries, human oversight, education, governance, and evidence of safe use. These findings are consistent with broader health care AI literature, which argues that the translation of AI into clinical practice depends on more than model performance and requires workflow integration, clinician trust, clear accountability, and prospective evaluations in real-world settings [58,59].

The mapping of the evidence demonstrated that the literature is currently unevenly distributed across professional groups, AI applications, and areas of clinician experience. Evidence was concentrated among psychiatrists, psychologists, and counselors or therapists, whereas comparatively few studies included allied health professionals, support workers, or other mental health professional groups. This imbalance is important because different professional groups have distinct roles, scopes of practice, patterns of patient contact, and responsibilities within mental health services, and may therefore experience the introduction of AI in different ways. Similarly, some AI applications, including symptom monitoring and prediction, documentation, and feedback or supervision tools, were examined across a narrower range of professional groups than general AI, GenAI, and conversational agents. The domain-level mapping further demonstrated comparatively limited evidence regarding clinicians’ AI knowledge and prior use and the impact of AI on therapeutic relationships. These gaps suggest that the existing literature may provide a stronger understanding of clinicians’ general attitudes toward AI, rather than of how specific technologies are experienced by the diverse workforce expected to use them in practice. Future research should therefore broaden representation across professional groups and examine specific AI applications within the clinical contexts in which they are intended to be implemented, with particular attention to clinicians’ real-world experience with these technologies and their effects on therapeutic care.

Successful implementation of AI into mental health care should begin with clinical need, rather than technological capability. Clinicians were most receptive to AI when it addressed current pressures on mental health services, such as administrative burden, workforce constraints, the abundance of information, and delays in patients’ access to care, suggesting that implementation should prioritize tools that address clearly defined problems within existing models of care. Introducing AI simply because the technology is available risks the introduction of tools that are poorly aligned to clinical workflows, increase burden, or fail to address issues within mental health practice, reflecting wider implementation concerns [58]. Lower-risk clinician-facing applications, such as documentation support, may provide a practical starting point to build clinicians’ knowledge and familiarity with AI tools, while preserving more time and attention for human interaction and clinical judgment [58-60].

To best support clinicians’ usage of AI, education should be treated as a prerequisite for successful implementation, rather than optional additions [61,62]. Clinicians require training that extends beyond surface-level familiarity to include the ability to generate prompts, critically appraise outputs, communicate with patients about appropriate AI use, understand privacy and confidentiality, recognize bias or hallucinations, manage overreliance, and be aware of accountability [63-65]. Moving forward, training programs, such as microcredentials, will be imperative for clinicians to develop core competencies required to safely use AI in mental health care. Legal responsibility and accountability also require careful consideration. Broader health care–AI literature has highlighted that existing liability frameworks may be poorly suited to address harms associated with AI, particularly where responsibility may be distributed across clinicians, health care organizations, software developers, service providers, and regulators [66,67]. Implementation will require clear governance, defining the intended roles of AI tools, clinicians’ responsibilities in checking outputs, documentation requirements, escalation pathways, and processes for monitoring errors or adverse events [59,61].

This review also highlights the need for co-design and real-world evaluation, as clinicians voiced a desire to be involved in AI implementation. Co-design is critical in mental health care as the consequences of poor implementation extend beyond technical errors and may affect patient safety, equity, therapeutic relationships, and trust [59,61,62,68]. This aligns with co-design literature in mental health care, which argues that interventions should be designed and evaluated with, rather than for, the people who will use them [68]. A tool that appears accurate in controlled testing environments may still not be successful if it is disruptive to therapeutic relationships, increases administrative burden, produces inaccurate outputs, or is perceived to be replacing human elements of care. AI development and implementation should be guided not only by clinicians but also by other key stakeholders including patients, carers, developers, researchers, and policymakers, to ensure that they are usable, acceptable, and clinically appropriate. Current AI evaluation guidance emphasizes the need to assess human-AI interactions, workflow integration, safety, and usability rather than technical performance alone [59,61,62,69].

As a scoping review, this study aims to map the existing literature rather than critically appraise study quality or establish causal relationships, which limits the ability to draw conclusions about the effectiveness or impact of AI-based tools in clinical practice. The review is also constrained by the scope and quality of the included studies, many of which rely on self-reported attitudes and perceptions that are susceptible to response bias, social desirability effects, and variations in AI literacy among clinicians. Additionally, the rapidly evolving nature of generative AI means that clinicians’ views captured in the literature may quickly become outdated as tools, governance frameworks, and clinical integration mature. Finally, heterogeneity in study designs, clinical settings, professional roles, and definitions of AI may limit the generalizability of the findings across different mental health contexts and health systems.

This review is innovative in shifting the focus from the technical performance of AI tools to the perspectives of the clinicians who will be expected to use, interpret, explain, and remain accountable for them in mental health care. Unlike previous reviews, this review provides a clinician-centered understanding of AI adoption, highlighting that acceptability depends not only on what AI can do but also on whether it can be integrated safely, ethically, and in ways that preserve professional judgment and therapeutic relationships. These findings suggest that implementation should begin with lower-risk (eg, administrative), clinician-facing applications, be supported by education and governance, and be evaluated in mental health settings before wider adoption.

Since the release of ChatGPT [55] in 2022, public awareness of AI-enabled tools has increased rapidly, prompting broader reflection on where these systems can and should be used. AI is rapidly becoming ubiquitous in health care, and as a result, mental health services will increasingly need to determine not only whether AI tools can be used but whether they can be implemented safely, ethically, and in ways that preserve clinicians’ judgment and patients’ trust. For this reason, future research should move to examine and evaluate how AI tools function in real-world mental health settings. This should include assessment of clinical workflow, clinician workload, patient experience, equity, privacy, safety, therapeutic relationships, and unintended consequences over time. Integration of AI into mental health care will therefore require ongoing evaluation and refinement, rather than one-off validation before implementation. The future of implementation into mental health care will therefore depend on ensuring that AI strengthens services without eroding human judgment, trust, and therapeutic relationships that remain central.