The scoping review was conducted in line with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines for conducting systematic scoping reviews (Checklist 1) [30]. The protocol for this scoping review was prospectively registered in Open Science Framework [31].

We included primary research that evaluated a ChatGPT application, implemented for a mental health care–specific purpose, and reported on a performance-related outcome. Performance-related outcomes were operationalized as any qualitative or quantitative data regarding, but not limited to, accuracy, precision, acceptability, feasibility, safety, usability, efficacy, strengths, or limitations of ChatGPT performing a specific task in the mental health care landscape. We focused only on clinical applications of ChatGPT, such as prediction, detection of mental health problems, psychological interventions, or clinical decision-making, while excluding studies investigating the use of ChatGPT for research, educational, technical, or administrative purposes. Reviews, as well as studies that describe the development of a ChatGPT application without reporting any performance-related outcomes, were excluded. Studies focusing solely on other generative AI technologies (eg, Claude, Copilot, and Gemini) were also excluded.

The first author conducted a search in April 2025 in multidisciplinary and specific domain databases (Web of Science, PubMed, Scopus, PsycINFO, Association for Computing Machinery Digital Library, IEEE Xplore, Open Access Theses and Dissertations, EBSCO, and ProQuest). A sample of the search strategy used is presented in Multimedia Appendix 1.

Screening of articles for inclusion was performed in 2 stages: title and abstract review and full article review, conducted independently by 2 reviewers. Following an initial screening of titles and abstracts, full texts were obtained and screened by 2 reviewers. Any divergences were solved through discussions between the 2 reviewers. The screening procedure was piloted under Cochrane guidelines, with a random sample of studies for both abstract and full text [32].

A standardized data extraction form was designed before data charting. The form was piloted and refined with the screening team. Similar to the study selection process, the 2 reviewers independently conducted the process of data extraction, with discrepancies being resolved by discussions and consensus.

From the included studies, relevant information was charted in an Excel (Microsoft) spreadsheet: (1) type of publication (peer-reviewed article, conference proceedings, working papers, etc), (2) purpose of application (detection/assessment, therapeutic application, decision making, and prognosis), (3) mental health problem focus, (4) age category of intended end users, (5) type of ChatGPT model (standard, custom instruction, custom GPT), (6) study design/methodology (prompt study, quasi-experimental, controlled study, study case, (7) participants, (8) comparison element (MH practitioners, other AI models), (9) outcomes assessed, and (10) the main findings. A detailed overview of the definitions for each item, along with its corresponding categories, is provided in Table S1 of the Multimedia Appendix 2.

Consistent with methodology for scoping review, data were synthesized using a descriptive and thematic approach [33]. We first conducted a numerical summary of study characteristics (eg, publication type, mental health focus, study design, ChatGPT version, etc). Then, we grouped findings by major application domains (detection, counseling/treatment, clinical decision support, and prognosis) following a deductive approach, where each study was assigned to the predetermined categories developed during the protocol stage. Finally, we presented a narrative synthesis of main findings to identify overarching patterns in performance, comparisons with mental health professionals or other AI systems, as well as variations across tasks and model versions, and evidence gaps. Regarding the relative performance of ChatGPT compared to mental health experts or other AI models, this reflects the comparative conclusions reported in individual studies, rather than a statistical synthesis across studies.

The detailed study selection process is presented in Figure 1, the PRISMA flowchart. A total of 4780 articles were identified in the search. After eliminating duplicates, 2342 abstracts were screened for title and abstract, with an additional 2149 articles being excluded. Out of the 193 remaining articles, 172 full-text copies were retrieved that were screened in full. This resulted in 60 articles being included in the current review. The detailed characteristics of the included studies are presented in Multimedia Appendix 3.

Summative results, as per characteristics of ChatGPT clinical applications and research, are detailed in Table 1. Most of the articles were published in peer-reviewed journals (n=47) [10-12,14-19,34-71], followed by conference proceedings (n=9) [13,36,72-78], and preprints (n=4) [79-82]. Regarding the purpose of the application, ChatGPT was predominantly employed as a tool for counseling and interventions in mental health care (n=29) [16-19,34-36,41,43,44,46,47,49,54-62,66-68,71,72,76,80], and detection/assessment of mental health problems (n=24) [10-13,15,15,42,48,51-53,63-65,69,70,73-75,77-79,81,82]. A few studies explored its application in supporting clinical decision-making (n=8) [14,15,37-39,42,50,52], respectively in prognosis (n=3) [15,40,83]. While a substantial portion of the studies addressed mental health in general (n=16) [11,16,17,34,35,37,38,43,53,55,58,59,71,72,76,80], others focused on specific conditions, including depression (n=15) [13,15,40,41,47,50,62,63,68,73,75,78,79,81,82], suicidality (n=12) [12,15,47,51,56,64,65,67,70,74,77,81], anxiety (n=8) [15,34,41,49,52,54,78,79], schizophrenia (n=4) [15,42,45,83], substance use disorders (n=3) [44,61,66], and autism spectrum disorders (n=3) [46,57,69]. Additionally, attention deficit hyperactivity disorder [18,36], and post-traumatic stress disorder [10,15] were each the primary focus of two studies (n=2), whereas individual studies addressed bipolar disorder [60], obsessive-compulsive disorder [48], insomnia [39], and self-harm [14].

The age category of the intended end users of the clinical applications consisted mostly of adults (n=56) [10-12,14-19,34-70], with only 4 studies evaluating the use of ChatGPT for detection, counseling, and clinical decision facilitation for mental health problems among children and adolescents [11,36,52,69]. Regarding the ChatGPT model specifications, most studies employed standard ChatGPT (n=50) [11-16,19,34,37-48,50-53,55-70,72,73,76-83]. Customized ChatGPTs were used in 6 studies [18,36,47,49,54,71], while 4 studies employed a custom instruction GPT model [10,63,74,75].

Most studies were designed as prompt experiments (n=50), in which the accuracy or the quality of ChatGPT-generated responses to various queries were evaluated, without involvement of human participants [10-18,36-48,50-53,55-57,59-67,69,70,72-83]. The designs of the remaining studies included uncontrolled clinical trials (n=5) [19,34,35,54,71], controlled trials (n=3) [49,58,68], and case reports (n=2) [42,43]. Only 10 studies enlisted participants to use/test ChatGPT as a part of an experimental setup. Among these 10 studies, adults from the general population were involved in 4 studies [43,49,68,71], while 6 other studies had participants from the clinical population [19,34,35,42,54,58]. The number of participants varied between 1 and 399. Participants were predominantly young adults with a high educational level. Dropout rates were generally low, except for 1 study that involved the elderly [68]. The performance in the specific clinical tasks conducted by ChatGPT was assessed by comparison with mental health practitioners (n=19) [11-13,15,35,38,40,45,46,50-52,56,58,59,66,68,69,83] or with other AI models (n=21) [13,15,17,36,39,40,44,46,48,49,52,53,56,62,67,74,76,77,79,82,83].

The main findings for ChatGPT use in detection, counseling and intervention, clinical decision facilitation, and prognosis of mental health care are presented in Multimedia Appendix 4.

Since its release in November 2022, ChatGPT has sparked extensive discussions in the mental health care sector [20,85]. However, its performance in conducting various clinical tasks has received less attention. This scoping review provides an insight into the clinical applications of ChatGPT in mental health care and its current empirical evidence.

The landscape of clinical use of ChatGPT is expanding, albeit unevenly, with a focus on detection, counseling, and treatment of a wide range of mental health problems, indicating the perceived value of ChatGPT to augment psychological services, especially where access is limited. However, its relatively infrequent use in areas requiring higher clinical accountability—such as prognosis and decision-making—suggests ongoing concerns about reliability, risk, and ethical responsibility [20]. Moreover, the widespread focus on standard ChatGPT, with minimal use of customized or fine-tuned models, represents a missed opportunity to strengthen context-sensitive adaptations critical for safe and effective clinical deployment [86]. Most clinical applications of ChatGPT in mental health care are primarily designed to be used for adults’ mental problems, with far fewer tools to benefit children and adolescents. This imbalance is striking, as these younger “Digital Natives” are often the earliest adopters of new technologies, and neglecting their needs risks creating a critical gap in safe, developmentally appropriate mental health support [87]. From a methodological stance, there is an overreliance on prompt-based experiment designs, based on simulations, without involving an interaction of real-world users. Even fewer studies involved clinical populations, which raises serious questions about whether ChatGPT is ready to be deployed at a large scale in mental health care services.

Although ChatGPT shows potential in conducting clinical tasks related to mental health care, research consistently fails to replicate the positive findings regarding performance. Besides the complexity of clinical tasks and presentations, another potential explanation for these inconsistencies might be related to the prompting and the level of pretraining used in the experimental testing [95]. Indeed, previous research has shown that the performance of ChatGPT in carrying out various tasks is highly dependent on the prompting engineering—namely, on how much task-specific information or training the model is given [96,97]. Several studies included in the current review have explicitly addressed this issue, showing, for example, that adding more examples in the prompt on how to carry out the detection tasks enhances the ChatGPT detection capabilities compared to zero-shot prompting condition, where ChatGPT relies purely on its pretrained knowledge to understand the task from the instructions users write in the prompt [10,63,74]. Similarly, use of the chain-of-thought technique improves diagnostic accuracy, since the model is encouraged to reason step-by-step—explicitly outlining its thought process—before arriving at a diagnostic or evaluative conclusion [63]. Additionally, a study showed that providing multimodal input, namely speech rhythm and rate, besides text-based data, increased ChatGPT’s accuracy in distinguishing between anxiety and depression [73]. In counseling and treatment, encouraging development is the growing evidence regarding the superiority of customized ChatGPTs, suggesting that specific domain optimization maximizes the benefits across the mental health domain, by addressing some of the limitations of generic AI models [18,36]. Another key moderator of ChatGPT’s performance in clinical practice is the model version, with newer iterations like GPT-4 generally outperforming GPT-3.5, though not consistently across all tasks. These results indicate that advances improve overall reliability but do not eliminate domain-specific weaknesses.

The findings of this review can serve as a guide to inform clinical practice regarding which type of ChatGPT applications and under which specific conditions can or cannot be reliably, safely, and confidently used, and which cannot. ChatGPT use should be limited to simple detection tasks such as binary decisions in initial screenings, triage, and continuous monitoring—if it examines or focuses on well-defined symptom constellations. It can also be used to manage and assist with counseling and intervention for simple and straightforward tasks and for simple clinical presentations, making it suitable for psychoeducation, low-intensity psychological treatments, and for support or cases where immediate care is not available. Within university counseling centers, such applications could help manage high service demand by providing first-line psychoeducational support and triaging students. In community mental health centers, ChatGPT could serve as a scalable adjunct to extend care to underserved populations, particularly in rural or low-resource contexts. In hospital-based or specialized clinical programs, its role may be more appropriately limited to intake assistance, between-session monitoring, or delivery of standardized interventions that complement provider-led care. However, given that the existing evidence with real-world patients and multicultural populations is scarce, implementation in these types of settings needs to be done with high caution. Additionally, our review suggests that ChatGPT in clinical practice should be regarded as merely a complementary tool and not a substitute for traditional mental health care, especially in complex or high-risk situations, where the value of human judgment and experience in decision-making is irreplaceable [41]. Additionally, when possible, users should choose fine-tuned or customized ChatGPT models over generic ones, because the former provide a higher level of sophistication and specificity [86,98]. While ChatGPT could be beneficial in assisting detection, counseling, and treatment, as well as in facilitating clinical decision-making for simple case presentations, both mental health experts and the clinical population should avoid turning to ChatGPT to forecast the trajectories of having mental health disorders, given its overly pessimistic outlook.

Several limitations of the current research must be noted. First, the inclusion of gray literature can pose issues regarding the quality of the study. However, in a fast-paced domain such as ChatGPT use, gray literature enhances comprehensiveness and timeliness of available evidence [99]. As this was a scoping review, we did not conduct a formal quality appraisal of included studies, consistent with Joanna Briggs Institute and PRISMA-ScR (Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews) guidance [30,33]. While the inclusion of gray literature broadened the scope of evidence, it also introduced variability in methodological rigor. Findings should therefore be interpreted with caution and regarded as exploratory, highlighting areas where more robust, peer-reviewed research is needed. Second, the methodology used preponderantly to test the performance of ChatGPT, namely prompt experiments, limits the conclusions regarding the ecological validity and how service users interpret or respond to AI outputs. Therefore, more rigorous testing designs are needed, including randomized controlled trials, exploring the additional benefits of using ChatGPT in traditional mental health care. Third, studies, including real-world users, are subject to demographic and self-selection biases, as they involve mostly young, highly educated adults who are likely to be more technologically literate and more open to digital tools, limiting generalizability.

Fourth, an important limitation emerges from the metrics used to assess ChatGPT’s performance. Accuracy or quality of answers to queries, as well as sophistication of conversation, do not equate with clinical efficacy and do not capture the process and mechanisms underlying its use, which are the main criteria for evidence-based practice in mental health care [100]. Therefore, future research should move beyond these metrics to assess whether ChatGPT use leads to symptom reduction and how it works. On the other hand, it cannot be asserted with certainty that the negative findings related to ChatGPT performance reflect actual AI deficits or that they are an artifact of distrust, negative perception, and attitudes of those who conducted the performance assessment. Algorithm aversion is a well-documented phenomenon in the AI field, referring to a default skepticism, a cognitive bias, where individuals distrust algorithm decisions and recommendations [101]. In mental health care, this aversion can lead practitioners and patients to favor human judgment over AI, even when AI demonstrates superior performance. For example, it has been shown that general trust in ChatGPT was a significant predictor of its perceived usefulness in clinical practice among health care practitioners [102]. Moreover, even the mere belief in AI involvement can diminish patients’ trust in medical and mental health-related advice, despite it being identical to that provided by human experts [103,104]. Addressing the main concerns related to trust, privacy, and ethics through education, transparent evaluation frameworks, and involving mental health care professionals in the development process is crucial for successfully adopting ChatGPT in mental health settings. Another significant issue in the use of ChatGPT for clinical applications in mental health care is related to the outdated training data it relies on. Most of the studies included tested ChatGPT 3.5 and 4, for which the cut-off date of training is September 2023; consequently, the clinical application does not integrate the latest developments. This aspect might be especially problematic in the mental health care domain, where clinical protocols for mental health disorder management are subject to ongoing updates, informed by new research findings [105].

Future research integrating ChatGPT in mental health clinical practice would also benefit from a multidisciplinary and coparticipatory approach. For example, given the encouraging results of fine-tuned and customized ChatGPT models, a further step would be an ongoing collaboration between AI and mental health experts in developing appropriate prompts for end users. Participatory methods provide one means of ensuring that AI-based solutions for mental health care are designed to meet users’ needs and therefore promote longer-term engagement [106]. The broader implications of deploying ChatGPT in mental health contexts must be addressed. The deployment of ChatGPT must be done within the existing and evolving regulatory and ethical frameworks [107]. A responsible integration of ChatGPT in mental health care involves built-in safeguarding mechanisms for accurate referrals, real-time escalation protocols for critical situations, and transparent accountability structures [107].

Future developments for ChatGPT in mental health care should prioritize training on domain-specific datasets (eg, psychiatric case notes, suicide risk assessments, and culturally diverse dialogues), and integration with evidence-based frameworks to enhance accuracy and therapeutic relevance [108]. Embedding established guidelines (Diagnostic and Statistical Manual of Mental Disorders, fifth edition, National Institute for Health and Care Excellence, and American Psychological Association recommendations) into model prompts or training and structured approaches such as CBT or acceptance and commitment therapy could make output more clinically reliable and standardized. Prognostic accuracy also requires improvement, through calibration with longitudinal clinical data, which could reduce the current negative bias [109]. Furthermore, enhancing cultural and contextual sensitivity through diverse training datasets would make the technology more equitable across populations [110].

In conclusion, this scoping review highlights the dual promise and perils of integrating ChatGPT into mental health care. While its scalability, immediacy, and overall diagnostic accuracy in categorical decisions and good therapeutic abilities make it a good candidate for addressing the need for immediate care, especially where the human workforce is not available, several limitations emphasize the need for cautious deployment in real life and clinical practice. The pitfalls include underperformance in complex and high-risk clinical situations, outputs lacking nuanced clinical reasoning and reliable references, and raising ethical and safety concerns. Consequently, at this moment, ChatGPT should be integrated as a supportive, not standalone, tool in mental health care, with careful oversight and adherence to ethical frameworks to ensure safety and effectiveness. Finally, we consider it crucial to address not only the inherent limitations of ChatGPT itself but also the general perception of users, particularly mental health practitioners, regarding the deployment of this tool in clinical practice. The default skepticism of users might contribute to the dismissal of this tool, ignoring its tremendous potential.