Ethics approval was obtained from the Queen’s University Health Sciences and Affiliated Teaching Hospitals Research Ethics Board (6041701). Participation was voluntary, anonymous, and uncompensated. Respondents provided informed consent before answering.

This cross-sectional mixed methods study examined MHPs’ perceptions of AI in clinical care via a structured online survey capturing quantitative and qualitative data. Guided by implementation science and ethical inquiry principles, it explored attitudes toward AI integration, perceived utility, and adoption-related challenges and facilitators.

Eligible participants were licensed or trainee MHPs in the United States. Prior AI experience was not required, but participants needed a basic understanding of AI in health care. A mixed nonprobability sampling strategy was used. Email outreach used convenience sampling by distributing the survey through professional networks and mailing lists affiliated with OPTT to reach individuals meeting our specific inclusion criteria. Social media recruitment via LinkedIn used convenience and snowball sampling. Recruitment took place from August 12, 2024, to November 18, 2024. This duration helped maintain stable contextual factors and allowed for multiple recruitment waves without risking major shifts in practice norms. Interested individuals reviewed a letter of information and provided implied consent via survey completion. Email reminders were sent monthly, and social media posts were made monthly.

The survey was developed by an interdisciplinary team of experts in digital mental health, psychiatry, and clinical psychology. We did not use validated psychometric instruments or a specific theoretical framework. Item development was instead informed by a literature review of the current landscape of AI in mental health. The items went through multiple rounds of review by the research team to ensure clarity and relevance. AI was intentionally left undefined to capture broad subjective interpretations of the technology as it currently exists in public discourse. The finalized 12-item instrument covered 6 domains: clinical background, AI familiarity and adoption intent, perceived applications, barriers and ethics, training and equity, and open-ended feedback. Response formats included multiple choice, Likert-type scales, check-all-that-apply items, and 1 open-ended item. The survey was internally pretested for clarity, usability, and face validity. The 6 domains covered by the finalized instrument were as follows:

The survey was administered via Qualtrics (Qualtrics International Inc) [13] optimized for desktop and mobile, with options to pause and resume. Reminders were sent via recruitment channels. No identifying information was collected. Data were stored on encrypted, password-protected servers accessible only to the research team.

Quantitative data were analyzed using the built-in Stats iQ module within the Qualtrics platform. To maintain transparency and reproducibility, statistical handling was defined. Missing data were managed via pairwise deletion, ensuring maximum retention of available data for each specific analysis without discarding entire cases. Survey items formatted as “select all that apply” were independently coded as mutually exclusive binary variables (selected vs unselected) for analysis. Before inferential testing, statistical assumptions were evaluated. For all chi-square tests of independence, expected cell frequencies were verified to ensure that they exceeded 5; where assumptions were met, standard unranked tests were run. Chi-square tests of independence were used to explore associations between AI use and other variables such as comfort levels, endorsement of clinical applications, and views on AI’s impact on care disparities. We evaluated effect size using the Cramér V and set significance at a P value below .05. No α-level adjustments were made for multiple comparisons. Consequently, these inferential analyses were exploratory and hypothesis generating.

Responses to the open-ended question were exported as a CSV file, and a descriptive qualitative summary approach was taken. One researcher reviewed all exported open-ended responses and conducted initial coding to organize the text into preliminary topical categories. Following this initial step, a second researcher reviewed the raw responses alongside the proposed categories. The 2 researchers discussed the coding structure and reached a consensus to finalize 6 overarching thematic categories.

A total of 62 licensed or trainee MHPs who were actively engaged in clinical care participated in the survey (Table 1). Most (n=35, 56.5%) held a master’s degree, whereas 33.9% (n=21) were PhD-level MHPs. In total, 6.5% (n=4) of the respondents primarily worked in administrative or leadership positions in mental health care but still maintained clinical responsibilities. Additional respondents included trainees (n=2, 3.2%).

Regarding current use of AI, most respondents (43/62, 69.4%) reported not using any form of AI in their clinical practice. When asked about anticipated integration timelines, the largest proportion of respondents (25/61, 41%) indicated that they were undecided. A smaller subset (13/61, 21.3%) had already begun incorporating AI tools into their practice. Other respondents planned to adopt AI in the next 6 (5/61, 8.2%), 7 to 23 (1/61, 1.6%), 24 to 35 (1/61, 1.6%), and ≥36 (1/61, 1.6%) months. Notably, nearly one-quarter (15/61, 24.6%) stated that they did not intend to adopt AI at all.

Participants were asked to evaluate various domains in which AI could be applied within mental health care (Table 2). The most widely endorsed application was for administrative functions such as coding and billing, with over half (32/60, 53.3%) of the respondents supporting this use case. Additional perceived applications of AI included intake and triage, automated documentation and progress note generation, decision support for diagnosis and treatment, chatbot-based patient support tools, and monitoring of behavioral or biological markers (Table 3). In the “Other” option of this survey item, several participants also identified unique or emergent areas where they believed AI could be beneficial, including patient education, marketing and outreach, scheduling, virtual care, clinical training, and protocol monitoring.

A chi-square test revealed that whether MHPs used AI in clinical practice significantly influenced their comfort using AI tools for assessment and treatment planning (n=61, χ²₄=10.2; P=.04; Cramér V=0.401). Among those using AI, 55.6% (10/18) reported feeling either somewhat or very comfortable with such tools compared to only 16.3% (7/43) of nonusers. Conversely, discomfort was more common among nonusers, with 67.4% (29/43) of participants reporting being either somewhat or very uncomfortable using AI. This suggests that AI users are more likely to express confidence in integrating AI into clinical workflows (Table 4).

When asked about respondents’ concerns about the use of AI in clinical practice, the most frequently cited were inaccurate outcomes (47/61, 77%) and algorithmic bias (41/61, 67.2%), followed by reduced accessibility for marginalized populations (31/61, 50.8%) and threats to job stability (22/61, 36.1%). Beyond these predefined options, respondents cited additional concerns in the “Other” option. MHPs voiced skepticism about the overuse of chatbots for therapeutic engagement, warning that excessive reliance on such tools could lead to the dehumanization of care. Several respondents highlighted the human connection in the therapeutic process and cautioned against treating AI as a substitute for relational work. Specific concerns included threats to patient privacy, regulatory breaches, and loss of emotional nuance in communication. Further comments suggested that AI tools were perceived as lacking sophistication, emotional intelligence, and linguistic precision. Several respondents also expressed concern about AI’s inability to comprehend complex human experiences, particularly in contexts involving trauma or relational dynamics. Ethical ambiguity surrounding the division of responsibility between MHPs and machines was another theme, as was the potential for cybersecurity breaches in sensitive clinical data environments. One seasoned MHP, reflecting on nearly 5 decades of experience, mentioned that the use of AI contradicts the foundational principles of therapeutic care. A large majority (54/61, 88.52%) rated formal training on the ethical and practical use of AI as important, underscoring the need for competency development before integration.

In response to whether AI could reduce disparities in access to mental health services, responses were mixed. While many participants acknowledged the potential for AI to improve access, some were less enthusiastic (Table 2). Overall, most MHPs felt that AI had the potential to reduce disparities to a “moderate” or “great” extent, although this optimism was tempered by recognition of the need for intentional, equitable implementation strategies. A chi-square test revealed a significant association between current use of AI in clinical practice and belief that AI improves access to mental health care across populations (n=61, χ²₄=14.0; P=.007; Cramér V=0.479). Among current users of AI, 61.1% (11/18) believed that AI improves access either moderately or to a great extent, whereas only 23.3% (10/43) of nonusers shared that view. Conversely, most nonusers felt that AI had no benefit, were unsure, or felt that it had a small benefit. These findings suggest that MHPs who engaged with AI were more likely to view it as a tool for improving equity in care delivery.

MHPs were asked what they felt were the most significant barriers to the adoption of AI in clinical care over the following 2 years. The most frequently selected challenge was MHP resistance to change or reluctance to modify established workflows (46/61, 75.4%). A closely related concern was uncertainty regarding the clinical validation and efficacy of AI tools (41/61, 67.2%), highlighting a hesitancy to integrate AI into practice without robust, peer-reviewed evidence of benefit. Technical integration challenges also featured prominently, with embedding AI systems into existing electronic health records and digital infrastructure being the most cited challenge among participants (26/61, 42.6%). Financial concerns were also mentioned (25/61, 41%) surrounding the cost of acquiring, maintaining, and updating AI systems in budget-constrained environments. Complexity or confusion surrounding the purchasing process was also mentioned (24/61, 39.3%), as was ambiguity regarding payer reimbursement structures for AI-supported care (24/61, 39.3%). Organizational readiness emerged as another constraint. Just over one-quarter of respondents (16/61, 26.2%) indicated that a lack of administrative support or leadership buy-in could pose a barrier to AI implementation.

In the “Other” option, several responses underscored the role of institutional culture in either facilitating or impeding innovation. Beyond structural and financial concerns, MHPs also expressed apprehension about patient acceptance. Some respondents noted that clients may be concerned about data privacy, cybersecurity, or surveillance when AI is used in clinical contexts. Others highlighted concerns about liability for incorrect AI-generated diagnoses or treatment recommendations. These concerns reflect a broader mistrust not only of the technology itself but also of the broader ecosystem into which it would be introduced. Respondents also pointed to the need for clearer delineation of clinical responsibility when AI tools are involved in decision-making. Several MHPs emphasized the importance of preserving therapeutic integrity, cautioning that overly rapid implementation without ethical safeguards, end user training, and patient education could risk doing more harm than good.

Effective AI integration will require relational sensitivity, ethical rigor, and cultural humility [23]. Co-design with MHPs, diverse validation studies, and transparent accountability structures are critical. MHP leadership is essential to safeguard therapeutic integrity, equity, and trust in digitally augmented care.

Limitations include a relatively small, self-selected sample and cross-sectional design, limiting generalizability across clinical populations and settings and introducing the risk of selection bias [24]. The quantitative analyses were exploratory in nature. Multiple comparisons were conducted without α-level correction, which increases the risk of type I errors. The inclusion of a single open-ended item, while allowing for some exploration of how and why respondents held certain views, limited the depth of insight that could be obtained. Future studies should use richer qualitative methods such as interviews or focus groups to explore these perspectives in greater depth. Furthermore, AI was intentionally left undefined in the survey to capture broad subjective interpretations of the technology as it currently exists in public discourse. While this choice allowed for a broader assessment of baseline attitudes, variations in how participants conceptualized AI constitute a limitation, as differing interpretations may have influenced specific ratings and comfort levels. Another limitation is that the survey was not formally pretested with a representative sample following subsequent post hoc revisions. Only the research team reviewed these later changes.

Future research should track how attitudes shift with exposure, training, and real-world use and examine impacts on clinical outcomes, professional identity, and patient engagement. Studies should also prioritize perspectives from marginalized providers and communities to ensure equitable AI development and deployment.

MHPs are neither wholly resistant nor uncritically accepting of AI in mental health care. Instead, they offer a balanced perspective that values administrative utility while fiercely protecting the relational depth of clinical practice. The integration of AI in psychiatry and psychotherapy is ultimately a human challenge rather than a technical one, deeply intertwined with professional identity and systemic equity. Because mental health care relies uniquely on the therapeutic alliance, implementation models from broader medicine cannot simply be copied and pasted into this field. Responsible innovation will require dedicated clinical co-design, transparent liability frameworks, and proactive training to ensure that digital tools support rather than compromise patient-centered care.