Artificial Intelligence in Clinical Supervision and Training of Psychologists and Other Mental Health Professionals: A Scoping Review.
Clinical psychology & psychotherapy · 2026
Abstract originale
BACKGROUND: Clinical supervision is fundamental to the development of competent mental health professionals; however, established approaches are limited by challenges in scalability, direct observation and timely feedback. Artificial intelligence has emerged as a potential adjunct to supervision and clinical training, yet the full scope of existing applications has not been systematically mapped. OBJECTIVE: This scoping review systematically mapped artificial intelligence applications in the clinical supervision and training of psychologists and other mental health professionals. METHODS: The review was conducted in accordance with Joanna Briggs Institute guidance and reported following the PRISMA-ScR. The protocol was registered on the Open Science Framework. Searches were conducted in nine databases, with supplementary searches in Google Scholar. Two reviewers independently screened records; charted data were checked against the original reports. RESULTS: Database searches identified 11,975 records; 4381 duplicates were removed, 7594 records were screened, 817 reports were assessed for eligibility and 73 sources of evidence were included. Four application categories were identified: virtual patient simulation, automated performance assessment, artificial intelligence-enhanced supervisory feedback and training analytics. Most studies were published from 2022 onward and were concentrated in structured therapies, university-based settings and high-income countries. CONCLUSIONS: Artificial intelligence is best positioned as an adjunct that expands opportunities for practice, observation and structured feedback, rather than as a replacement for human supervision. Future research should employ more rigorous designs, validated competency measures, transparent reporting of artificial intelligence systems and outcomes relevant to trainees, supervisors and clients.
Parole chiave
Leggi il testo integrale originale
Testo estratto dalla pubblicazione: sono esclusi immagini, media, supplementi e parti soggette a diritti distinti.
Artificial Intelligence in Clinical Supervision and Training of Psychologists and Other Mental Health Professionals: A Scoping Review
ABSTRACT
Background
Clinical supervision is fundamental to the development of competent mental health professionals; however, established approaches are limited by challenges in scalability, direct observation and timely feedback. Artificial intelligence has emerged as a potential adjunct to supervision and clinical training, yet the full scope of existing applications has not been systematically mapped.
Objective
This scoping review systematically mapped artificial intelligence applications in the clinical supervision and training of psychologists and other mental health professionals.
Methods
The review was conducted in accordance with Joanna Briggs Institute guidance and reported following the PRISMA‐ScR. The protocol was registered on the Open Science Framework. Searches were conducted in nine databases, with supplementary searches in Google Scholar. Two reviewers independently screened records; charted data were checked against the original reports.
Results
Database searches identified 11,975 records; 4381 duplicates were removed, 7594 records were screened, 817 reports were assessed for eligibility and 73 sources of evidence were included. Four application categories were identified: virtual patient simulation, automated performance assessment, artificial intelligence‐enhanced supervisory feedback and training analytics. Most studies were published from 2022 onward and were concentrated in structured therapies, university‐based settings and high‐income countries.
Conclusions
Artificial intelligence is best positioned as an adjunct that expands opportunities for practice, observation and structured feedback, rather than as a replacement for human supervision. Future research should employ more rigorous designs, validated competency measures, transparent reporting of artificial intelligence systems and outcomes relevant to trainees, supervisors and clients.
Summary
Current AI applications in clinical training are concentrated in simulated clients, automated competency assessment, supervisory feedback and training analytics.
AI may support practice and structured feedback, but increased self‐efficacy and satisfaction should not be interpreted as demonstrated clinical competence.
Human supervisors remain responsible for case conceptualization, supervisee anxiety and distress, transference and countertransference, clinical risk and ethical and cultural judgement.
Training programs implementing AI should establish governance structures for privacy, bias monitoring, data security, trainee support and escalation procedures to human supervisors.
Evaluate observed therapeutic performance and transfer to clinical practice separately from confidence, diagnostic recognition and the accuracy of AI‐generated scores.
1Introduction
Mental health systems internationally are under sustained strain. The Global Burden of Disease Study estimates that depression and anxiety disorders increased by more than 25% during the early COVID‐19 pandemic (World Health Organization 2022). Demand for qualified mental health professionals consistently surpasses supply, particularly in low‐ and middle‐income countries, where the treatment gap for common mental disorders exceeds 75% (Hanna et al. 2018; Hoeft et al. 2018).
Clinical supervision supports trainees in developing competence, integrating evidence‐based practices into clinical work and managing the demands of practice (Bradley and Becker 2021; Choy‐Brown and Stanhope 2018; Iosim et al. 2021). Its effects on client outcomes remain uncertain: a recent meta‐analysis found a small, statistically non‐significant pooled effect on patient symptoms (Schreyer et al. 2025). Reviewing session recordings and assessing treatment fidelity can require substantial staff time and expertise (Marriott et al. 2026). When supervision relies on supervisees' accounts alone, aspects of clinical practice may remain unobserved and trainees may withhold relevant information (Choy‐Brown and Stanhope 2018; Foskett and Van Vliet 2021). In low‐resource settings, limited access to psychotherapy‐trained supervisors creates additional barriers to structured supervision (Rimal et al. 2025).
These limitations have substantial practical consequences. In routine community mental health care, supervision time may be available while direct observation and other feedback activities recommended in best‐practice models remain less consistently reported (Choy‐Brown and Stanhope 2018). Monitoring adherence to evidence‐based protocols is difficult when reviewing and coding session recordings requires time and trained staff (Marriott et al. 2026). Generative AI has been proposed as a way to supplement psychiatric education, including simulation and assessment, but evidence of educational effectiveness remains limited (Lee et al. 2025).
1.1Artificial Intelligence in Mental Health Education
Artificial intelligence (AI) has been explored for diagnostic support, monitoring and psychoeducation in mental health (Olawade et al. 2024). Applications relevant to training include automated evaluation of psychotherapy material (Hirsch et al. 2018), fidelity feedback (Creed et al. 2022) and structured practice of counselling responses (Tanana et al. 2019). Large language models (LLMs) also enable simulated client conversations (Holderried et al. 2026; Sawah et al. 2026). These applications differ in their proximity to supervision: some evaluate an algorithm or a simulated task, whereas others are embedded in clinical training. Technical feasibility alone does not establish educational effectiveness.
1.2Limitations of Generative AI and LLMs
Generative AI introduces limitations that are directly relevant to clinical training. LLMs generate text from patterns learned during training; fluent output does not establish factual accuracy or clinical understanding (Bender et al. 2021). Training applications therefore require evaluation of both their outputs and the consequences of using those outputs in learning.
Incorrect or clinically inappropriate feedback could reinforce errors in practice. Bias is also a concern: Bouguettaya and Aboujaoude (2025) found racial differences in psychiatric recommendations generated by LLMs under otherwise comparable case conditions. Finally, conversational fluency can encourage users to attribute relational capacities to a system that has no professional responsibility for the interaction. Ethical analyses have raised concerns about artificial empathy, attachment, privacy and accountability in mental health applications (Fiske et al. 2019). These issues warrant human review and explicit investigation in training settings.
1.3Prior Reviews and the Knowledge Gap
Previous reviews provide complementary perspectives. Magill et al. (2022) mapped technology‐based counselling skills training before the widespread availability of LLMs. Aafjes‐van Doorn et al. (2021) examined machine learning across psychotherapy research and Laricheva et al. (2026) examined NLP applications in counselling and psychotherapy. More recent reviews addressed AI in psychiatric education and in psychiatry and psychology education (Lee et al. 2025; Prégent et al. 2025; Weightman et al. 2025). The present review focuses on applications relevant to supervision and professional training, with particular attention to what their reported outcomes can establish about learning and clinical competence.
This review addresses these gaps by systematically mapping AI applications in the supervision and training of psychologists and other mental health professionals, with particular focus on the period following the widespread availability of LLM‐based tools. A preliminary search of PROSPERO, PubMed, the Cochrane Database of Systematic Reviews and JBI Evidence Synthesis confirmed the absence of current or in‐progress reviews on this specific topic. The review therefore asks: What are the existing applications of AI in the clinical supervision and training of psychologists and other mental health professionals? The objectives were to identify the range of AI applications in supervision and training contexts, map relevant training contexts and psychological domains, characterize AI technologies and reported functionalities and identify gaps calling for further investigation.
2Methods
2.1Study Design
Given that this review aimed to identify and map the available literature, a scoping review methodology was considered most appropriate (Munn et al. 2018). The review was conducted in accordance with methodological guidance for scoping reviews established by the Joanna Briggs Institute (JBI) and reported following the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses extension for Scoping Reviews (PRISMA‐ScR).
2.2Protocol and Registration
The protocol for this review was registered on the Open Science Framework platform (https://osf.io/h8473/) to ensure methodological transparency. The results were reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses extension for Scoping Reviews (PRISMA‐ScR), including a flow diagram detailing the study selection process (Tricco et al. 2018).
2.3Research Question
The central research question, structured using the PCC acronym, was: What are the existing applications of AI in the clinical supervision and training of psychologists and other mental health professionals?
Population (P): psychologists in clinical supervision, supervisors and other mental health professionals in training.
Concept (C): Applications of AI (e.g., machine learning, natural language processing, chatbots, and generative language models) in clinical training and supervision.
Context (C): Implementation in scholarly or occupational clinical training settings.
2.4Eligibility Criteria
2.4.1Participants
Studies were eligible if they included psychologists in clinical supervision, supervisors responsible for professional development, or other mental health professionals in training. This included psychologists at initial or intermediate stages of clinical practice, residents or interns in specialization or multidisciplinary residency programs, counsellors in training and supervisors who monitored and evaluated trainee progress.
2.4.2Concept
The review encompassed the set of AI applications, including machine learning techniques, natural language processing (NLP), chatbots and generative language models, aimed at improving clinical training and supervision. Particular emphasis was placed on the use of these technologies for the analysis of therapeutic processes, the simulation of clinical interactions and the provision of structured feedback to foster competence and expertise in professionals in training.
2.4.3Context
The review was limited to practices of implementation and evaluation of these technologies in academic and clinical professional training settings. This context encompassed undergraduate courses, postgraduate programs, residencies, teaching clinics and other supervised or institutional settings aimed at training and developing skills in mental health.
2.4.4Types of Sources
This scoping review included studies containing experimental and quasi‐experimental designs (randomized controlled trials, non‐RCTs, before‐and‐after studies and interrupted time‐series studies), analytical observational studies (prospective and retrospective cohorts, case–control studies, analytical cross‐sectional studies) and descriptive observational designs (case series, individual case reports and descriptive cross‐sectional studies). Studies using quantitative, qualitative, or mixed‐methods approaches were considered. Grey literature included reports, theses, conference proceedings, guidelines, policies, websites and opinion pieces. Reference lists of identified scoping and systematic reviews were screened to identify further relevant studies.
Studies focused exclusively on patients without an interface with supervision or professional training were excluded to ensure methodological focus.
2.5Search Strategy
The search strategy was outlined in three sequential steps as recommended by Peters et al. (2021). Preliminary research was conducted to support the introduction and justification of the protocol, drawing on the prior expertise of one author. On March 21, 2025, using the advanced search function, an initial search was carried out in Ovid/MEDLINE, as detailed in Supporting Information 1. This supplementary appendix provides the full Ovid/MEDLINE search strategy, including search terms, operators and record counts. As the screening process took longer than planned, an update was conducted on May 22, 2026, covering the period from March 21, 2025, to May 22, 2026, to identify additional eligible publications.
Subsequently, the strategy was adapted and replicated across eight additional databases (PsycINFO, EMBASE, LILACS, Web of Science, Scopus, ERIC, IEEE Xplore and ACM Digital Library) to ensure the retrieval of multidisciplinary evidence and methodological harmony across heterogeneous platforms.
Grey literature was considered an important source for scoping reviews, especially in quickly developing areas such as AI. Google Scholar was limited to the first 100 results. Reference lists of all included studies were screened using the snowball method. No language or time restrictions were applied. DeepL Translator and Google Translate were used to assess non‐English sources for eligibility; sources that could not be translated were excluded.
For eligible studies with incomplete data or methodological omissions, a protocol was established to contact corresponding authors via email. Articles whose authors did not respond within 10 business days were excluded to ensure the validity of the evidence synthesis.
2.6Study Selection
After the search was completed, all identified articles were imported into the Rayyan software (Ouzzani et al. 2016) for organization and screening. A pilot test was conducted in which two reviewers evaluated 25 articles based on the eligibility criteria to guarantee consistency. Discrepancies had been discussed and the screening process continued once a minimum agreement level of 75% was achieved.
Following the pilot test, the remaining titles and abstracts were independently and blindly reviewed by two reviewers. Articles deemed potentially relevant were retrieved for full‐text assessment. Disagreements at any stage were resolved through discussion; when unresolved, a third reviewer made the final decision. The PRISMA‐ScR checklist (Tricco et al. 2018) guided the reporting process.
2.7Data Charting
Data charting was performed using a structured instrument aligned with the PCC mnemonic and the review question. The instrument, based on JBI guidance for scoping reviews, was adapted to meet the specific requirements of the study. AI‐assisted preliminary extraction was performed in Rayyan (Ouzzani et al. 2016) only to support initial charting; each extracted field was verified and compared with the original source by human reviewers.
The data extraction protocol incorporated the following predefined categories: authors, year of publication, country of origin, study objective, population and sample size, study methodology, type of AI used, main findings and conclusions or recommendations.
2.8Data Analysis and Presentation
Data analysis included frequency counts for descriptive quantitative synthesis and high‐level thematic categorization of AI applications, training contexts, outcomes and implementation issues. Results are presented using a PRISMA‐ScR flow diagram, a summary table and narrative synthesis aligned with the review objectives. Consistent with scoping review methodology, no formal critical appraisal was undertaken; therefore, findings are interpreted as a map of the evidence base rather than as conclusive evidence of effectiveness.
3Results
3.1Study Selection
Searches identified 11,975 records. After removal of 4381 duplicates, 7594 records were screened by title and abstract and 6777 were excluded. A total of 817 reports proceeded to full‐text assessment; 744 were excluded and 73 sources of evidence were included in the synthesis. Conference publication, dissertation status and review design were not exclusion criteria in themselves. Figure 1 presents the study selection process.
3.2Characteristics of Included Sources
The mapping included 73 sources of evidence relevant to AI applications in clinical supervision and training, spanning 2008 to 2026. The characteristics and full bibliographic details of all 73 included sources are provided in Supplementary Table S1. The table also identifies sources included in the descriptive mapping that are not discussed individually in the narrative synthesis. Publications were concentrated in recent years and encompassed empirical evaluations, methodological studies and other eligible sources. This range provides a broad account of research activity but should not be interpreted as 73 independent tests of training effectiveness.
Empirical designs included randomized comparisons, before‐and‐after evaluations, qualitative studies and computational validation. Protocols and reviews contributed information about planned research and the wider literature. Sample sizes and units of analysis varied, including trainees, other participants, sessions, utterances and datasets; these units should not be combined as if they represented equivalent training samples.
Technologies included GPT‐family LLMs, custom machine learning models and specialized clinical platforms such as Lyssn. Settings included university programs, residency training, community services and online training. Applications addressed CBT, motivational interviewing, general counselling skills, suicide prevention and multicultural orientation. Some evidence was indirect because the evaluated population, task, or dataset differed from routine supervision of mental health professionals. Table 1 summarizes the four application categories, predominant AI technologies and typical functions and settings.
TABLE 1Summary of included studies by application category.
Application category Predominant AI technology Typical functions and settings
Virtual patient simulation LLMs, chatbots, conversational agents, voice‐based avatars Low‐stakes clinical role‐play, repeated skills practice, realistic client narratives and debriefing in university, residency and online training settings.
Automated performance assessment Custom machine learning models, NLP classifiers, transformer models Classification of therapist utterances, coding of discrete therapeutic skills, fidelity assessment, empathy ratings and multicultural‐orientation indicators.
AI‐enhanced supervisory feedback LLMs and specialized clinical supervision platforms Post‐session summaries, live or structured feedback, alternative clinical interpretations, technique recommendations and support for human supervision.
Training analytics and outcome monitoring NLP, topic modelling, prediction models, working‐alliance analytics Analysis of training/session data, outcome prediction, monitoring of therapeutic discourse and large‐scale quality improvement in training datasets.
Note: Categories are descriptive groupings derived from high‐level content categorization rather than mutually exclusive intervention classes.
Abbreviations: AI = artificial intelligence; LLM = large language model; NLP = natural language processing.
3.3Reported Training and Clinical Outcomes
The reported outcomes addressed different questions and should be interpreted separately. Self‐efficacy and perceived improvement concern learners' appraisals; performance on a simulated task concerns behaviour in that task; algorithm validation concerns the quality of a scoring or classification system; and client outcomes concern clinical change. None can automatically substitute for another.
For behaviour‐based, automatically scored task performance, Tanana et al. (2019) randomized 151 nontherapist participants to conditions in a ClientBot study. The proportion of utterances classified as reflections was 91% higher with feedback and 76% higher after feedback was removed than in the comparison group. These are between‐group differences in a specific counselling behaviour, not percentage gains in overall competence; the scoring used automated classifiers. Young et al. (2026) reported greater improvement in AI‐scored CBT skills among trainees receiving AI feedback, using historical cohorts and analogue role‐plays. The nonrandomized design and use of AI for both feedback and assessment limit conclusions about independently assessed competence and transfer to clinical practice.
3.3.1Algorithm Validation and Diagnostic Recognition
Algorithm‐validation studies reported agreement with human ratings on selected tasks. Lubbe et al. (2026) reported a correlation of R = 0.60 for empathy ratings in a segment‐based assessment approach. Hammerfald et al. (2025) reported an F1 score of 0.78 and Cohen's κ of 0.69 for identifying microcounseling skills. These estimates describe scoring or classification performance within the study conditions; they do not demonstrate equivalence to the full role of a human supervisor or show that trainees improved.
In a feasibility study with AI‐simulated patients, medical students correctly identified depression severity in 60% and suicidality phase in 67% of cases (Holderried et al. 2026). These descriptive recognition outcomes do not establish improvement from baseline, appropriate treatment selection, or the ability to deliver effective treatment. Goldberg et al. (2025) reported correlations between machine‐generated and human ratings of multicultural orientation; these results concern measurement of recorded material rather than the development of trainees' cultural competence.
3.3.2Client Outcomes in Training Settings
Sheperis et al. (2025) described reductions in depression and anxiety symptoms in a university clinic using AI‐supported supervision. Such observations are clinically relevant, but do not isolate the contribution of AI from treatment, human supervision, or other features of the service. Kuo, Tanana, et al. (2024) analyzed the prediction of client distress from session recordings; this is evidence about outcome monitoring rather than a test of an AI training intervention. Patient improvement in a dataset should not be presented as evidence that an AI tool caused better care.
3.3.3Affective and Confidence Outcomes
Jeong et al. (2025) reported increases in counselling self‐efficacy and reductions in anxiety in a comparison of AI feedback with self‐review following chatbot practice. Although self‐efficacy trajectories differed, post hoc between‐group contrasts were not significant after multiplicity correction. Steenstra et al. (2025), Levkovich et al. (2025) and Elyoseph et al. (2025) also reported self‐efficacy outcomes in their respective training contexts. Sanz et al. (2025) measured negative affect and perceived knowledge and skills. These findings concern confidence, affect and perceived capability. They do not by themselves demonstrate acquisition of therapeutic skills. Grundmann et al. (2025) reported a decline in self‐efficacy after use of a conversational agent, indicating that confidence may change in either direction.
3.3.4Performance Feedback Quality
Feedback quality and perceived usefulness were assessed separately from trainee competence. Herschbach et al. (2025) reported that 85.38% of real‐time feedback was partially or totally correct; 83.64% of participants considered it accurate and 87.27% found it helpful. A judgement of partial correctness is not equivalent to consistently correct clinical guidance. Zech et al. (2022) evaluated prediction of human‐rated facilitative interpersonal skills, while Pellemans et al. (2024) evaluated automated coding of motivational interviewing behaviours. These studies support investigation of assessment tools but do not establish that receiving their scores improves treatment skills.
3.3.5Comparisons Between Training Formats
Comparative findings depended on the outcome assessed. Sawah et al. (2026) reported more favourable participant ratings for voice‐based avatars than text‐based chatbots, including perceived skill improvement. Cioffi et al. (2025) compared ratings of supervisory feedback generated under different conditions, including human feedback. Perceived improvement and favourable ratings of a feedback product should be distinguished from independent demonstrations of trainee skill development. These comparisons do not establish the general superiority of AI‐supported supervision.
3.4Categories of AI Implementation
3.4.1Virtual Patient Simulation
Virtual patient simulation was a prominent application, enabling repeated practice through text or voice interactions. Studies examined the plausibility of generated client narratives, usability and stability of the simulated role (Tian et al. 2025; Cabrera Lozoya et al. 2025; Holderried et al. 2026). These are feasibility and experience outcomes; demonstrating that a simulation is usable or realistic is an initial step toward evaluating its educational value.
Systems varied in interactivity and realism. Text‐based chatbots and voice‐based avatars offered different experiences of practice (Sawah et al. 2026). The FASSLING approach combined role‐play with coaching and debriefing (Zhu 2025). Participants in other studies described overly agreeable responses, limited complexity and insufficient nonverbal information (Schmidt et al. 2025; Akkurt et al. 2025). Such limitations may constrain practice of challenging relational situations.
3.4.2Automated Performance Assessment and Feedback
Automated assessment applications classified therapeutic behaviours or estimated ratings from clinical material. Hammerfald et al. (2025) examined microcounseling skills and Lubbe et al. (2026) compared segment‐based and holistic assessment procedures. Agreement with human raters was task‐dependent. Standardized output should not be assumed to be unbiased or to capture the contextual judgement required in supervision.
Herschbach et al. (2025) examined use of specific communication techniques alongside real‐time feedback. Imel et al. (2024) evaluated detection of suicide risk assessment in crisis‐counselling calls, reporting different performance at call and statement levels. Detection of whether an assessment occurred is distinct from evaluating its adequacy, clinical decision‐making, or the effectiveness of the subsequent intervention.
3.4.3Supervisory Support and Professional Development
Applications also supported reflection on clinical material. Matthews et al. (2025) examined AI as a reflective partner and Sokolovskaya (2024) discussed supervisory uses such as generating alternative interpretations and resources. Hsieh et al. (2024) evaluated expert judgments of ChatGPT‐generated case conceptualizations. Mendlovic (2025) described AI‐assisted deliberate practice in psychodynamic psychotherapy. These reports show that higher‐order functions have received exploratory attention, but evaluation of generated content is different from evidence that trainees acquired these competencies.
AI‐supported clinic workflows included session summaries and monitoring of interventions, themes and symptom change (Sheperis et al. 2025). Peavy et al. (2024) discussed technology‐supported counselling and supervision in opioid treatment programs. Such functions may help supervisors select material for review; their clinical value depends on verification of the outputs and integration into human supervisory work.
3.4.4Training Data Analytics and Outcome Monitoring
Kuo, Tanana, et al. (2024) predicted next‐session symptom ratings from the linguistic content of the previous session (test‐set Spearman's ρ = 0.32, p < 0.001; R
2 = 0.07). These are predictive‐performance estimates rather than intervention effects. Topic modelling was also used to describe patterns in therapeutic discourse (Vanin et al. 2025).
The COMPASS framework used language modelling to examine patient‐therapist alliance strategies (Lin et al. 2024). Computational representations of an alliance provide potential material for investigation and feedback, but are not direct demonstrations of improved alliance management or training effectiveness.
3.4.5Ethical and Cultural Issues Across Applications
Ethical and cultural issues occurred across the four application categories. İçen and Köken (2025) examined AI‐generated guidance in ethically challenging child and adolescent psychiatry scenarios. Kuo, Mehta, et al. (2024) examined automated identification of cultural conversations in therapy. Producing guidance or recognizing cultural content is not equivalent to demonstrating that a trainee can reason ethically or respond appropriately to a client's cultural context.
Studies of simulated clients also raised questions about cultural representation and the effects of prompting (Akkurt et al. 2025; García‐Torres et al. 2025). These findings support attention to culturally responsive design and supervisory review of model outputs, rather than assuming that a plausible simulation adequately represents diverse clients.
3.5Patterns Across Study Contexts
The following patterns provide context for interpreting the mapping. Because outcomes, populations and designs differ, they should be treated as descriptive observations and hypotheses for further study rather than tested moderators of effectiveness.
3.5.1Modality and Interaction Design
Interaction format influenced users' experience: avatar and text‐based simulations received different ratings (Sawah et al. 2026). Assessment procedure also mattered, with segment‐based and holistic prompting producing different agreement with human ratings (Lubbe et al. 2026). These findings address distinct outcomes and cannot be combined into a single estimate of educational effectiveness.
3.5.2Feedback Timing and Structure
Feedback during practice was associated with more reflections in the ClientBot experiment (Tanana et al. 2019). In contrast, the comparison of AI feedback and self‐review in Jeong et al. (2025) concerned self‐efficacy and anxiety. Together these examples indicate why the content and timing of feedback must be considered alongside the type of outcome, rather than assuming that favorable confidence scores demonstrate skill acquisition.
3.5.3Training Stage and Prior Experience
Participants ranged from people without therapist training to professional trainees and practicing counsellors. Tanana et al. (2019) studied nontherapists, whereas Young et al. (2026) studied psychotherapy trainees. Differences between these studies cannot isolate the effect of the training stage. Transfer to supervised clinical work requires direct evaluation in the relevant population.
3.5.4Domain‐Specific Considerations
CBT and motivational interviewing offered structured behaviors and fidelity frameworks for automated assessment (Creed et al. 2022; Steenstra et al. 2025). Their representation in the literature does not establish greater AI effectiveness in these modalities. In particular, the Project AFFECT publication is a study protocol, not a completed effectiveness trial (Creed et al. 2022). Relational and culturally responsive performance requires evaluation beyond the detection of protocol‐consistent utterances.
3.5.5Implementation Context and Support
Training applications were embedded in varied educational and service contexts (Kopelovich et al. 2025; Schmidt et al. 2025). Human supervision, curricular activities and practice opportunities can accompany AI use and may contribute to observed change. Comparisons across these settings do not isolate the added value of AI or establish that a hybrid model is superior without an appropriate comparison design.
3.5.6Quality and Sophistication of AI Systems
Ratings of AI‐generated content varied with model configuration, prompting and the task used for evaluation (Cioffi et al. 2025; Akkurt et al. 2025; Shusterman et al. 2025). A favourable rating of a generated response is evidence about that response under the tested conditions. It should not be generalized to the system's ability to supervise a clinician or produce durable improvements in care.
4Discussion
4.1Principal Findings
This scoping review mapped 73 sources of evidence relevant to AI applications in clinical supervision and training. The four overlapping categories describe uses of technology, rather than four established forms of effective supervision. The literature includes simulations, tools for coding clinical material, systems that generate feedback and analyses of training or clinical‐session data. The distinction between these functions is necessary because evidence that a tool can perform a task is not necessarily evidence that its use improves professional competence.
The adjunctive interpretation follows from the outcome‐specific reading of the evidence. Practice studies show that selected behaviours can change in a simulated task (Tanana et al. 2019); self‐efficacy studies address perceived capability (Jeong et al. 2025); and validation studies address agreement with human judgements (Lubbe et al. 2026). These findings identify potentially useful components of training but leave a gap between the evaluated task and broad clinical competence. Reports of client change do not close that gap when they cannot isolate the contribution of AI.
Accordingly, the recommendation to retain human supervision is an interpretive conclusion based on the limited scope of the demonstrated outcomes and the wider functions of supervision. It is not a claim that hybrid supervision has been shown to be superior in all settings. The same reasoning informs the research priorities: independent performance assessment is needed to evaluate competence, clinical follow‐up to evaluate transfer and direct study of relational supervisory processes to evaluate higher‐order functions.
4.2Relationship to Prior Reviews
The present mapping complements reviews of technology‐based counselling training, machine learning in psychotherapy and AI in psychiatric and psychological education (Aafjes‐van Doorn et al. 2021; Magill et al. 2022; Lee et al. 2025; Prégent et al. 2025; Weightman et al. 2025). Its contribution is to consider these applications in relation to supervisory functions and the inferential distance between a reported outcome and competent clinical practice. This perspective allows feasibility, perceived benefit, algorithm performance and clinical learning to be discussed without treating them as interchangeable.
4.3Implications for Clinical Training and Supervision
Training programs should specify the competency being practised, the evidence needed to assess it and how a supervisor will review AI‐generated feedback. A trainee can report feeling more capable while still needing support with formulation, intervention choice, timing and responsiveness to the client. Conversely, reduced confidence may reflect recognition of complexity rather than reduced competence. Neither interpretation should be accepted without observation of performance.
The distinction is supported by research on therapist self‐assessment outside the included AI literature. In a cognitive processing therapy implementation trial, self‐efficacy increased but was unrelated to independently rated adherence and competence (Pace et al. 2021). Brosan et al. (2008) found overestimation of competence, especially among less competent therapists, whereas McManus et al. (2012) found a tendency toward underestimation in trainees. The direction of error can vary; limited experience alone should not be treated as proof of overconfidence. Confidence remains a meaningful educational outcome, but it requires calibration against observed performance.
Higher‐order supervision includes helping trainees construct and revise case conceptualizations, manage anxiety and distress and examine transference, countertransference and the supervisory relationship (Borders 2014; American Psychological Association 2025). These functions connect discrete techniques to a particular client, therapist and therapeutic process. The clinical value of AI therefore depends partly on whether its outputs help human supervisors examine these processes.
For case conceptualization, an AI‐generated formulation may provide hypotheses for discussion. Hsieh et al. (2024) evaluated the acceptability of generated conceptualizations, which does not by itself show that trainees can integrate developmental history, context, maintaining processes and treatment priorities in their own work. Supervisory review should ask trainees to justify and revise a formulation using case evidence and to connect it to intervention choices. Identifying a disorder or risk state is only one element of this reasoning; it is not equivalent to treating the condition.
For anxiety and distress, the relevant question extends beyond whether a simulation makes trainees feel more comfortable. Supervision must help them recognize how distress affects attention, judgement, avoidance, or clinical boundaries. The anxiety outcomes reported by Jeong et al. (2025) concern training‐related affect and do not establish that an AI system can manage sustained distress or its effects on clinical practice. AI‐generated prompts may assist reflection, while the supervisor evaluates the trainee's needs and the implications for safe practice.
For transference and countertransference, AI may suggest questions about recurring relational patterns or identify material for joint review. Such suggestions remain hypotheses: the supervisor and trainee must examine the interaction, emotional responses, context and competing explanations. Exploratory reflective and psychodynamic applications (Matthews et al. 2025; Mendlovic 2025) do not establish that AI can reliably recognize or work through these processes. Future studies should assess formulation quality, responses to relational ruptures, reflective capacity and clinically appropriate use of countertransference information, alongside discrete technical skills.
4.4Equity, Implementation and Governance
The concentration of studies in university‐affiliated and higher‐resource settings limits transferability to services with different languages, resources and training needs. Performance observed in a particular dataset or simulation may not generalize across cultures, professions, or clinical populations. Local evaluation should examine whose experiences are represented and whether a system's errors fall disproportionately on particular groups.
Implementation should therefore be guided by clinical governance rather than by technological availability alone. Training programs need explicit procedures for privacy, data security, bias monitoring, human review of AI‐generated feedback, trainee support and escalation when AI‐supported training reveals clinical risk or uncertainty.
4.5Strength of the Evidence Base
No formal risk‐of‐bias appraisal was undertaken. The following qualifications therefore describe the design and measurement issues visible in the mapping, rather than a graded assessment of evidence certainty. Small or short‐term studies, nonrandomized comparisons and self‐reported outcomes limit causal interpretation. Independent ratings, blinded assessment where feasible and follow‐up in actual clinical practice, are needed to assess durable skill acquisition. If a system both trains and scores the trainee, improvement may partly reflect alignment with that system's scoring criteria; external validation is essential. Clinical outcome analyses also need to account for human supervision, treatment exposure, case mix and clustering of clients within therapists.
4.6Limitations
This review has limitations at the level of the review process and the evidence mapped. First, it was designed to describe a heterogeneous literature and no formal critical appraisal or meta‐analysis was conducted. Application categories can overlap and narrative examples cannot establish pooled effectiveness, rank competing systems, or determine which training contexts benefit most. The interpretive distinctions between outcome types improve clarity but do not remove biases in the underlying studies.
Second, search and access decisions may have affected coverage. Restricting Google Scholar screening to the first 100 results, excluding sources that could not be translated and excluding reports with unresolved missing information may introduce retrieval or selection bias. AI‐assisted preliminary extraction in Rayyan also requires careful human verification because generated summaries can misclassify designs or outcomes. Rapid changes in models and their availability further limit the stability of a mapping of this field.
Third, short follow‐up, small or setting‐specific samples, self‐reported outcomes and limited independent assessment constrain inferences about competence and transfer. Evidence from simulations, algorithm‐validation datasets and groups outside routine mental health professional training is indirect for clinical supervision. Where AI both supplies feedback and assesses performance, shared scoring criteria may inflate apparent learning. Confidence, diagnostic recognition and ratings of generated content do not establish effective treatment delivery. Finally, client outcomes and higher‐order supervisory processes receive insufficient direct evaluation to support conclusions about patient benefit or replacement of human supervision. These limitations require cautious interpretation and define priorities for further research.
5Conclusion
This scoping review maps AI applications in clinical supervision and training across virtual patient simulation, automated assessment, supervisory feedback and training analytics. The evidence supports further investigation of these tools as adjuncts for selected learning and assessment tasks. It does not establish that increased self‐efficacy, successful diagnostic recognition, or agreement between AI and human ratings translates into broader treatment competence. Case conceptualization, management of supervisee anxiety and distress and exploration of transference and countertransference remain central to supervision and require explicit evaluation in AI‐supported models. Conclusions about client benefit and replacement of human supervisory functions are premature. Future research should combine independently assessed competence with clinical follow‐up, transparent model reporting and attention to cultural and service contexts. Overall, the findings support investigating AI as an adjunct to, rather than a replacement for, human clinical supervision. Its contribution should ultimately be judged not by the sophistication or efficiency of the technology alone, but by its demonstrated capacity to support the development of competent practitioners and the delivery of safe, effective and equitable mental health care.
Funding
The authors have nothing to report.
Conflicts of Interest
The authors declare no conflicts of interest.
Supporting information
References
Aafjes‐van Doorn , K. , C. Kamsteeg , J. Bate , and M. Aafjes . 2021 . “A Scoping Review of Machine Learning in Psychotherapy Research .” Psychotherapy Research
31 , no. 1 : 92 –116 . 10.1080/10503307.2020.1808729 .32862761
Akkurt , M. , R. Maurya , and T. Brown . 2025 . “Learning Through Simulation: Counselor Trainees' Interactions With ChatGPT as a Client .” Behavioral Science
15 , no. 12 : 1660. 10.3390/bs15121660 .PMC12730025 41464004
American Psychological Association
. 2025 . “Guidelines for Clinical Supervision in Health Service Psychology .” https://www.apa.org/about/policy/guidelines‐supervision‐revised.pdf.10.1037/a0038112 25581007
Bender , E. M. , T. Gebru , A. McMillan‐Major , and S. Shmitchell . 2021 . “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610‐623. 10.1145/3442188.3445922 .
Borders , L. D. 2014 . “Best Practices in Clinical Supervision: Another Step in Delineating Effective Supervision Practice .” American Journal of Psychotherapy
68 , no. 2 : 151 –162 . 10.1176/appi.psychotherapy.2014.68.2.151 .25122982
Bouguettaya , A. , and E. Aboujaoude . 2025 . “Racial Bias in AI‐Mediated Psychiatric Diagnosis and Treatment: A Qualitative Comparison of Four Large Language Models .” npj Digital Medicine
8 : 332 . 10.1038/s41746-025-01746-4 .PMC12137607 40467886
Bradley , W. J. , and K. D. Becker . 2021 . “Clinical Supervision of Mental Health Services: A Systematic Review of Supervision Characteristics and Practices Associated With Formative and Restorative Outcomes .” Clinical Supervisor
40 , no. 1 : 88 –111 . 10.1080/07325223.2021.1904312 .PMC8146512 34045790
Brosan , L. , S. Reynolds , and R. G. Moore . 2008 . “Self Evaluation of Cognitive Therapy Performance: Do Therapists Know How Competent They Are? ” Behavioural and Cognitive Psychotherapy
36 , no. 5 : 581 –587 . 10.1017/S1352465808004438 .
Cabrera Lozoya , D. , M. Conway , E. Sebastiano De Duro , and S. D'Alfonso . 2025 . “Leveraging Large Language Models for Simulated Psychotherapy Client Interactions: Development and Usability Study of Client101 .” JMIR Medical Education
11 : e68056. 10.2196/68056 .PMC12312989 40743466
Choy‐Brown , M. , and V. Stanhope . 2018 . “The Availability of Supervision in Routine Mental Health Care .” Clinical Social Work Journal
46 , no. 4 : 271 –280 . 10.1007/s10615-018-0687-0 .PMC6426317 30906079
Cioffi , V. , O. Ragozzino , L. L. Mosca , et al. 2025 . “Can AI Technologies Support Clinical Supervision? Assessing the Potential of ChatGPT .” Inform
12 . 10.3390/informatics12010029 .
Creed , T. A. , L. Salama , R. A. Slevin , et al. 2022 . “Enhancing the Quality of Cognitive Behavioral Therapy in Community Mental Health Through Artificial Intelligence Generated Fidelity Feedback (Project AFFECT): A Study Protocol .” BMC Health Services Research
22 : 1177 . 10.1186/s12913-022-08519-9 .PMC9487132 36127689
Elyoseph , Z. , I. Levkovitch , Y. Haber , et al. 2025 . “Using GenAI to Train Mental Health Professionals in Suicide Risk Assessment .” Journal of Clinical Psychiatry . 10.4088/jcp.24m15525 .40608482
Fiske , A. , P. Henningsen , and A. Buyx . 2019 . “Your Robot Therapist Will See You Now: Ethical Implications of Embodied Artificial Intelligence in Psychiatry, Psychology, and Psychotherapy .” Journal of Medical Internet Research
21 , no. 5 : e13216. 10.2196/13216 .PMC6532335 31094356
Foskett , A. J. , and K. J. Van Vliet . 2021 . “Understanding Supervisee Nondisclosures in Supervision With Videorecording Review and Interpersonal Process Recall .” Counselling and Psychotherapy Research
21 , no. 1 : 188 –197 . 10.1002/capr.12306 .
García‐Torres , D. , C. Fernández , J. J. Mira , et al. 2025 . “PsicoSimGPT: Evaluating the Use of Generative AI for Training in Psychopathological Interviewing .” 10.2196/preprints.78857 .
Goldberg , S. B. , M. J. Tanana , S. H. Stewart , et al. 2025 . “Automating the Assessment of Multicultural Orientation Through Machine Learning and Natural Language Processing .” Psychotherapy (Chicago, Ill.)
62 : 292 –300 . 10.1037/pst0000519 .PMC11291704 38300571
Grundmann , S. , M. Al Owayyed , M. Bruijnes , et al. 2025 . “Lilobot: A Cognitive Conversational Agent to Train Counsellors at Children's Helplines: Design and Initial Evaluation .” Journal of Medical Systems
49 : 5 . 10.1007/s10916-024-02121-8 .PMC11729055 39806193
Hammerfald , K. , S. Schmidt , V. Vlassov , F. Schmidt , H. Haaland Jahren , and O. A. Solbakken . 2025 . “Leveraging Large Language Models to Identify Microcounseling Skills in Psychotherapy Transcripts .” Psychotherapy Research
36 : 1058 –1076 . 10.1080/10503307.2025.2539405 .40817802
Hanna , F. , C. Barbui , T. Dua , A. Lora , M. van Regteren Altena , and S. Saxena . 2018 . “Global Mental Health: How Are We Doing? ” World Psychiatry
17 , no. 3 : 367 –368 . 10.1002/wps.20572 .PMC6127814 30192105
Herschbach , L. , T. Festl‐Wietek , C. Stegemann‐Philipps , et al. 2025 . “Evaluation of an AI‐Based Chatbot Providing Real‐Time Feedback in Communication Training for Mental Health Care Professionals: Proof‐Of‐Concept Observational Study .” Journal of Medical Internet Research
27 : e82818. 10.2196/82818 .PMC12701347 41314643
Hirsch , T. , C. S. Soma , K. Merced , et al. 2018 . ““It's Hard to Argue With a Computer”: Investigating Psychotherapists' Attitudes Towards Automated Evaluation .” Designing Interactive Systems. 10.1145/3196709.3196776 .PMC6050022 30027158
Hoeft , T. J. , J. C. Fortney , V. Patel , and J. Unützer . 2018 . “Task Sharing Approaches to Improve Mental Health Care in Rural and Other Low‐Resource Settings: A Systematic Review .” Journal of Rural Health
34 , no. 1 : 48 –62 . 10.1111/jrh.12229 .PMC5509535 28084667
Holderried , F. , A. Sonanini , A. Philipps , et al. 2026 . “Using AI to Train Future Clinicians in Depression Assessment: Feasibility Study .” JMIR Medical Education
12 : e87102. 10.2196/87102 .PMC12946781 41678789
Hsieh , L.‐H. , W.‐H. Liao , and E.‐Y. Liu . 2024 . “Feasibility Assessment of Using ChatGPT for Training Case Conceptualization Skills in Psychological Counseling .” Computers in Human Behavior: Artificial Humans
2 . 10.1016/j.chbah.2024.100083 .
İçen , S. , and A. H. Köken . 2025 . “Artificial Intelligence Guidance in Ethically Challenging Clinical Scenarios in Child and Adolescent Psychiatry: A Qualitative Study in the Context of Türkiye .” BMC Medical Ethics
26 : 183. 10.1186/s12910-025-01323-0 .PMC12754976 41469990
Imel , Z. E. , B. Pace , B. Pendergraft , et al. 2024 . “Machine Learning‐Based Evaluation of Suicide Risk Assessment in Crisis Counseling Calls .” Psychiatric Services
75 : 1068 –1074 . 10.1176/appi.ps.20230648 .PMC11530329 39026467
Iosim , I. , P. Runcan , V. Dan , B. Nadolu , R. Runcan , and M. Petrescu . 2021 . “The Role of Supervision in Preventing Burnout Among Professionals Working With People in Difficulty .” International Journal of Environmental Research and Public Health
19 , no. 1 : 160. 10.3390/ijerph19010160 .PMC8750667 35010424
Jeong , D. , H. Lee , H. Yang , J. C. Park , and S. M. Lee . 2025 . “AI Integration in Counseling Training: Aiding Counselors‐In‐Training in Self‐Efficacy Enhancement and Anxiety Reduction .” Journal of Counseling and Development
103 : 346 –360 . 10.1002/jcad.12560 .
Kopelovich , S. L. , R. A. Slevin , R. Brian , et al. 2025 . “Preliminary Investigation of an Artificial Intelligence‐Based Cognitive Behavioral Therapy Training Tool .” Psychotherapy
62 : 12 –21 . 10.1037/pst0000550 .PMC11913585 39869688
Kuo , P. , M. Mehta , H. Hashtpari , et al. 2024 . “Identification of Cultural Conversations in Therapy Using Natural Language Processing Models .” Psychotherapy
61 : 259 –268 . 10.1037/pst0000542 .39404819
Kuo , P. , M. J. Tanana , S. B. Goldberg , et al. 2024 . “Machine‐Learning‐Based Prediction of Client Distress From Session Recordings .” Clinical Psychological Science
12 : 435 –446 . 10.1177/21677026231172694 .PMC11299859 39104662
Laricheva , M. , Y. Liu , E. Shi , and A. Wu . 2026 . “Scoping Review on Natural Language Processing Applications in Counselling and Psychotherapy .” British Journal of Psychology
117 , no. 2 : 677 –701 . 10.1111/bjop.12721 .PMC13051009 39095975
Lee , Q. Y. , M. Chen , C. W. Ong , and C. S. H. Ho . 2025 . “The Role of Generative Artificial Intelligence in Psychiatric Education: A Scoping Review .” BMC Medical Education
25 : 438. 10.1186/s12909-025-07026-9 .PMC11938615 40133891
Levkovich , I. , Y. Haber , Y. Levi‐Belz , and Z. Elyoseph . 2025 . “A Step Toward the Future? Evaluating GenAI QPR Simulation Training for Mental Health Gatekeepers .” Frontiers in Medicine
12 . 10.3389/fmed.2025.1599900 .PMC12187690 40568211
Lin , B. , D. Bouneffouf , Y. Landa , et al. 2024 . “COMPASS: Computational Mapping of Patient‐Therapist Alliance Strategies With Language Modeling .” arXiv. 10.48550/arXiv.2402.14701 .PMC12081631 40374613
Lubbe , D. , N. Wedderhoff , and C. Nelles . 2026 . “AI‐Driven Versus Human Evaluations of Psychotherapeutic Communication Performance: The Impact of Procedural Prompting on Assessment Reliability and Validity .” Computers in Human Behavior Reports
21 : 100910. 10.1016/j.chbr.2025.100910 .
Magill , M. , N. R. Mastroleo , and S. Martino . 2022 . “Technology‐Based Methods for Training Counseling Skills in Behavioral Health: A Scoping Review .” Journal of Technology in Behavioral Science
7 : 325 –336 . 10.1007/s41347-022-00252-8 .PMC8983031 35402697
Marriott , B. R. , A. E. Meyer , A. Feagans , B. L. Batch , and Z. W. Adams . 2026 . “Characterizing Fidelity Monitoring Practices in Community Behavioral Health Care .” Journal of Behavioral Health Services & Research
53 : 45 –64 . 10.1007/s11414-025-09967-x .PMC12876097 40931317
Matthews , A. , A. Lerman , M. Beach , et al. 2025 . ““It's Like Having That Supervisor in the Room”: Examining AI as a Reflective Partner .” Psychotherapy Research
36 : 1306 –1317 . 10.1080/10503307.2025.2569047 .41084310
McManus , F. , S. Rakovshik , H. Kennerley , M. Fennell , and D. Westbrook . 2012 . “An Investigation of the Accuracy of Therapists' Self‐Assessment of Cognitive‐Behaviour Therapy Skills .” British Journal of Clinical Psychology
51 , no. 3 : 292 –306 . 10.1111/j.2044-8260.2011.02028.x .22803936
Mendlovic , S. 2025 . “Using an AI‐Based System for Dimensional Deliberate Practice in Psychodynamic Psychotherapy .” BRAIN. Broad Research in Artificial Intelligence and Neuroscience
16 : 449 –463 . 10.70594/brain/16.3/34 .
Munn , Z. , M. D. J. Peters , C. Stern , C. Tufanaru , A. McArthur , and E. Aromataris . 2018 . “Systematic Review or Scoping Review? Guidance for Authors When Choosing Between a Systematic or Scoping Review Approach .” BMC Medical Research Methodology
18 : 143 . 10.1186/s12874-018-0611-x .PMC6245623 30453902
Olawade , D. B. , O. Z. Wada , A. Odetayo , A. C. David‐Olawade , F. Asaolu , and J. Eberhardt . 2024 . “Enhancing Mental Health With Artificial Intelligence: Current Trends and Future Prospects .” Journal of Medicine, Surgery, and Public Health
3 : 100099. 10.1016/j.glmedi.2024.100099 .
Ouzzani , M. , H. Hammady , Z. Fedorowicz , and A. Elmagarmid . 2016 . “Rayyan—A Web and Mobile App for Systematic Reviews .” Systematic Reviews
5 : 210 . 10.1186/s13643-016-0384-4 .PMC5139140 27919275
Pace , B. T. , J. Song , M. K. Suvak , N. Shields , C. M. Monson , and S. W. Stirman . 2021 . “Therapist Self‐Efficacy in Delivering Cognitive Processing Therapy in a Randomized Controlled Implementation Trial .” Cognitive and Behavioral Practice
28 , no. 3 : 327 –335 . 10.1016/j.cbpra.2020.08.002 .
Peavy , K. M. , A. Klipsch , C. S. Soma , et al. 2024 . “Improving the Quality of Counseling and Clinical Supervision in Opioid Treatment Programs: How Can Technology Help? ” Addiction Science & Clinical Practice
19 : 8 . 10.1186/s13722-024-00435-z .PMC10799386 38245783
Pellemans , M. , S. Salmi , S. Merelle , et al. 2024 . “Automated Behavioral Coding to Enhance the Effectiveness of Motivational Interviewing in a Chat‐Based Suicide Prevention Helpline: Secondary Analysis of a Clinical Trial .” Journal of Medical Internet Research
26 : e53562. 10.2196/53562 .PMC11327631 39088244
Peters , M. D. J. , C. Marnie , A. C. Tricco , et al. 2021 . “Updated Methodological Guidance for the Conduct of Scoping Reviews .” JBI Evidence Implementation
19 , no. 1 : 3 –10 . 10.1097/XEB.0000000000000277 .33570328
Prégent , J. , V. H. A. Chung , I. El Adib , M. Désilets , and A. Hudon . 2025 . “Applications of Artificial Intelligence in Psychiatry and Psychology Education: Scoping Review .” JMIR Medical Education
11 : e75238. 10.2196/75238 .PMC12340458 40720804
Rimal , P. , S. Shrestha , R. Khatri , et al. 2025 . “Addressing Challenges for Psychotherapy Supervision in Global Mental Health Through Experiential Learnings From Rural Nepal .” Discover Public Health
22 : 246. 10.1186/s12982-025-00645-z .PMC12064468 40352469
Sanz , A. , J. L. Tapia , E. García‐Carpintero , J. F. Rocabado , and L. M. Pedrajas . 2025 . “ChatGPT Simulated Patient: Use in Clinical Training in Psychology .” Psicothema
37 : 23 –32 . 10.70478/psicothema.2025.37.21 .40549860
Sawah , V. E. , A. Bhardwaj , A. Pryke‐Hobbes , et al. 2026 . “Artificial Intelligence as a Training Tool in Clinical Psychology: A Comparison of Text‐Based and Avatar Simulations .” arXiv. 10.48550/arXiv.2601.11533 .
Schmidt , S. , D. C. Lozoya , L. Kiropoulos , D. Cabrera Lozoya , M. Conway , and S. D'Alfonso . 2025 . “Psychology Student and Mental Health Practitioner Experiences of and Perspectives on Client101, a Virtual Client Chatbot Training Tool .” BMC Medical Education
25 : 1293 . 10.1186/s12909-025-07668-9 .PMC12492891 41039348
Schreyer , B. , C. Leithner , R. Eilers , K. Gossmann , and R. Rosner . 2025 . “The Effects of Clinical Supervision on Supervisees and Patient Outcomes in Psychotherapy: A Systematic Review and Meta‐Analysis .” Frontiers in Psychiatry
16 : 1705578. 10.3389/fpsyt.2025.1705578 .PMC12679832 41356684
Sheperis , D. , S. Sadeh‐Sharvit , V. Guinn , R. Joseph , E. Bunge , and L. Assenheim‐Dagan . 2025 . “Integrating Artificial Intelligence in a University Counseling Clinic .” Journal of Counselor Preparation and Supervision
19 , no. 3 : 1 –13 . 10.70013/wdo9cs4m .
Shusterman , R. , A. C. Waters , S. O'Neill , M. Bangs , P. Luu , and D. M. Tucker . 2025 . “An Active Inference Strategy for Prompting Reliable Responses From Large Language Models in Medical Practice .” npj Digital Medicine
8 : 119 . 10.1038/s41746-025-01516-2 .PMC11847020 39987335
Sokolovskaya , M. 2024 . “Enhancing Supervisory Practice in Positive and Transcultural Psychotherapy Through Artificial Intelligence .” Global Psychotherapist
4 : 223 –227 . 10.52982/lkj253 .
Steenstra , I. , F. Nouraei , and T. W. Bickmore . 2025 . “Scaffolding Empathy: Training Counselors With Simulated Patients and Utterance‐Level Performance Visualizations .” CHI Conference on Human Factors in Computing Systems. 10.1145/3706598.3714014 .
Tanana , M. J. , C. S. Soma , V. Srikumar , D. C. Atkins , and Z. E. Imel . 2019 . “Development and Evaluation of ClientBot: Patient‐Like Conversational Agent to Train Basic Counseling Skills .” Journal of Medical Internet Research
21 : e12529. 10.2196/12529 .PMC6662153 31309929
Tian , X. , Y. Liu , S. Litt , K. Li Zhang , K. Shin , and T. Jiang . 2025 . “Do Artificial Intelligence Clients Speak Like Human Clients? Exploring GPT‐4's Content‐Level Performance in Counseling Role‐Play .” Journal of Counseling and Development
104 : 14 –26 . 10.1002/jcad.70016 .
Tricco , A. C. , E. Lillie , W. Zarin , et al. 2018 . “PRISMA Extension for Scoping Reviews (PRISMA‐ScR): Checklist and Explanation .” Annals of Internal Medicine
169 , no. 7 : 467 –473 . 10.7326/M18-0850 .30178033
Vanin , A. , V. Bolshev , and A. Panfilova . 2025 . “Psychotherapist Remarks' ML Classifier: Insights From LLM and Topic Modeling Application .” Frontiers in Psychiatry
16 . 10.3389/fpsyt.2025.1608163 .PMC12332746 40787679
Weightman , M. J. , A. Chur‐Hansen , and S. R. Clark . 2025 . “AI in Psychiatric Education and Training From 2016 to 2024: Scoping Review of Trends .” JMIR Medical Education
11 : e81517. 10.2196/81517 .PMC12755346 41474738
World Health Organization
. 2022 . “Mental Health and COVID‐19: Early Evidence of the Pandemic's Impact: Scientific Brief, 2 March 2022 .” https://www.who.int/publications/i/item/WHO‐2019‐nCov‐Sci_Brief‐mental_health‐2022.1.
Young , J. , M. C. Schruff , E. Mudryk , and J. P. Bentley . 2026 . “Artificial Intelligence‐Generated Feedback as a Training Tool for Enhancing Psychotherapy Skill .” Journal of Technology in Behavioral Science . 10.1007/s41347-026-00603-9 .
Zech , J. M. , R. Steele , V. K. Foley , and T. D. Hull . 2022 . “Automatic Rating of Therapist Facilitative Interpersonal Skills in Text: A Natural Language Processing Application .” Frontiers in Digital Health
4 . 10.3389/fdgth.2022.917918 .PMC9425293 36052318
Zhu , Y. 2025 . “Simulation‐Based Clinical Training With AI: Integrating FASSLING for Enhanced Emotional Intelligence and Therapeutic Competency in Clinical Psychology Education .” 10.54254/3049-5458/2025.21247 .
Come citare questo studio
da Silva JVN, Ferreira PVS, Dos Santos PPS, Penido MA. (2026). Artificial Intelligence in Clinical Supervision and Training of Psychologists and Other Mental Health Professionals: A Scoping Review.. Clinical psychology & psychotherapy. DOI: 10.1002/cpp.70359
Consulta la pubblicazione originale ↗