Inclusivity and Innovation: A Pilot Study on Free AI Models for Clinical Supervision in Gestalt Therapy (GT)

1. Introduction

Clinical supervision is crucial for the practice of psychotherapy and for training in it: it is a cornerstone of professional support, stimulating clinical reflection and the continuous development of skills [1,2].

The supervisor, therefore, takes on the role of “agent of transformation,” with the main purpose of promoting the therapist's growth [3]. In Gestalt Therapy (GT) in particular, supervision takes on a unique form, based on phenomenology and the complexity of the individual-environment interaction [4,5,6].

GT pays particular attention to the experiential field, highlighting how individuality emerges from the interaction between organism and environment through processes of creative adaptation. Founded by Perls, Hefferline, and Goodman (1951), GT is based on a holistic understanding of the individual in the “here and now” and on the centrality of awareness in the therapeutic process [7]. In this context, supervision aims to explore the dynamics of interaction between therapist, patient, and field [8,9], and only secondarily to evaluate the technical effectiveness of therapeutic work [1,10,11].With the advent of advanced technologies and their increasingly rapid and surprising development, new frontiers of study continue to open up in the field of psychotherapy and all its areas [12,13,14,15,16,17,18]. Large Language Models (LLMs), including ChatGPT, have shown significant capabilities in analyzing, understanding, and producing human language, even when tested with reports of therapeutic processes [19,20,21]. In particular, recent studies explore the use of ChatGPT specifically as a supervision support tool: the results have highlighted its ability to support and integrate critical reflection and the identification of relevant points in therapy sessions [19,21]. It is clear that supervision, like psychotherapy itself, is characterized by great relational complexity and the need for a personalized approach [22]; inevitably, the question arises of how to communicate with LLM models such as ChatGPT to make them a useful tool for these practices [23,24].Cioffi et al. (2024) [19] have contributed to outlining the crucial importance of prompt design in optimizing the effectiveness of AI in supporting clinical supervision. The results of this study highlight that the appropriate and accurate formulation of the prompt is crucial for obtaining responses that are relevant, practically useful, and technically sound. In this case, a good prompt must be a clear and comprehensive conceptual framework and not just a list of instructions. A carefully constructed prompt can significantly transform the outcomes of interaction with AI, making them more precise, articulate, and useful in stimulating the therapist to reflect on their practice. The study showed: 1) that the best accuracy of supervisory feedback is achieved in the context of a preliminary interactive dialogue aimed at questioning the chatbot on topics of interest, correcting its inaccuracies, and testing it in generating dialogues related to therapy sessions and supervision, outlining the personal and professional characteristics of the ‘protagonists’ involved. 2) That the responses obtained from the chatbot in point 1 can be specifically synthesized by the chatbot itself into a single text, which effectively serves as a prompt-framework effectively replacing the progressive and detailed introduction of the topic referred to in point 1.

In a further development of this study [21], the authors further explored the potential of AI in supervision. The most representative feedback generated in the previous study was compared with feedback written by an expert human supervisor, using a Likert scale satisfaction questionnaire completed by a group of psychotherapy trainees. The subsequent statistical analysis of the responses to the satisfaction questionnaire (PCA, principal component analysis) revealed four dimensions: Relational and Emotional, which includes an empathetic approach, emotional impact, support for reflection, and reinforcement of the therapist's confidence; Didactic and Technical Quality, which includes clarity, relevance, completeness, and analysis of therapeutic techniques; Treatment Support and Development, which assesses the presence of practical suggestions, highlighting areas for improvement, and equal dialogue between colleagues; Professional Orientation and Adaptability, which includes orientation towards ethical principles, support for the therapeutic contract, and adequacy to the therapist's professional level.

The evidence emerging from the two studies cited, with regard to prompt design and the main components, forms the basis of the present study.

2. Aims

This study aims to contribute to exploring the use of artificial intelligence in the supervision of psychotherapists, with a particular focus on Gestalt Therapy. The article follows on from previous work [19,21] and aims to consolidate the basis for a more effective use of LLMs.

This study has two main objectives: one is to develop a more effective and functional prompt-framework than the one already tested, drawing on the evidence gathered in the two previous studies in the context of the current state of development of the chatbot, which is different from when it was previously used. The prompt tested in the first study [19] will be restructured to be more detailed, with the aim of providing a more effective frame, and will be integrated with the fundamental components that characterize quality feedback that emerged in the second study [21].

The prompt will therefore be enriched with elements that guide the machine to provide responses characterized by improved perceived empathy, ethical relevance, communicative clarity, and practical support.

The second objective of this study is to test, using the new prompt, the free ChatGPT-4o model, used through a free account created from scratch and without LTM (long-term memory) function, comparing its performance with ChatGPT-4o with LTM, used through a more ‘expert’ account, already used for previous studies and regularly used by a Gestalt therapist, and with the performance of ChatGPT-4, also used through the latter ‘expert’ account. The focus is on evaluating the ability of these models to generate feedback that meets the technical, relational, and practical needs of the therapist. In addition to understanding the qualitative differences between the models, the aim is to evaluate the potential of freely accessible tools, with a view to inclusivity towards therapists or trainees with limited resources.

3. Metodology

The methodology of this study is structured as follows: in the first phase, the old prompt was reformulated and enriched with ChatGPT-4 using the “expert” account, already used for previous studies and enriched by Long-Term Memory. This account was chosen for this phase because of its ability to integrate the experience accumulated in handling complex requests in areas of interest, having ‘learned’ from previous interactions. The request made to ChatGPT 4 was to enrich the previous prompt to make it more effective in adequately replacing the progressive introduction to the topic in providing the framework in which to insert the supervision request. The request also specified that the new wording of the prompt should also be suitable for use with free models on new accounts.

Once the new version of the prompt was obtained, it was integrated with information relating to the four main dimensions. The information integrated into the prompt at this stage is as follows: Relational and emotional dimension: it must be characterized by a linguistic structure that can have the effect of an empathetic approach on the reader; it must stimulate self-reflection for the therapist; it must stimulate the therapist's self-confidence in managing the clinical case; it must have an emotional impact.

"Didactic and technical quality: it must be formulated in a clear and understandable manner; it must be relevant to the clinical case presented; it must be complete and comprehensive with respect to the request; it must include an accurate analysis of the techniques used.

Treatment support and development: it must be helpful to the treatment; it must contain practical suggestions for the therapist; it must constructively highlight areas for improvement; it must appear as a communication between peers.

Professional orientation and adaptability: it must be ethically oriented; it must be helpful to the therapeutic contract; it must be appropriate to the therapist's professional level; it must be useful for the therapist's professional development."

Once this integration into the first form of the new prompt had been carried out, we asked ChatGPT-4 to reformulate the new text integrated with the new information, without changing its content, in order to ensure its linguistic adequacy for good interaction with the AI.

Finally, the final wording of the prompt was supplemented with an explicit request to the model, in the form of an open-ended question, to ask further questions to deepen the information needed to generate appropriate feedback. This addition aims to promote greater interactivity and maximize the AI's performance in understanding the therapeutic context. This is an essential element for effective supervision. Once the prompt was constructed, it was tested on three different configurations: ChatGPT-4 on an expert account with LTM function; ChatGPT-4o on an expert account with LTM function; ChatGPT-4o on a free and new account (first use) without LTM, simulating a context accessible to all.For each of the three chat sessions, after entering the prompt, the clinical case presented was the same as that used in previous studies (where the therapist is one of the co-authors of both this article and the previous works); however, the presentation of the case was preliminarily reformulated in order to make it more fluid and concise. In all three tests, the presentation of the case was followed by a question to the AI to verify whether further information was useful, so as to allow the machine access to all relevant data in order to obtain feedback that was as accurate and personalized as possible.

The test results were then analyzed to assess the ability of the different models to provide feedback consistent with the main components, responding to the specific requirements of Gestalt supervision.

The three supervisory feedbacks obtained were compared in relation to all the sub-dimensions of the four main components, integrating a semantic analysis carried out using ChatGPT with a thorough review and correction of the same by the researchers. The most subjective aspects of impact were evaluated by the therapist responsible for the case study.

4. Results

Table 1 shows that the supervisory feedback generated by the ChatGPT-4 model is characterized by a balance between emotional support, practical guidance, and technical analysis. The language used is warm and welcoming to read, helping to create the idea of a safe and functional supervisory environment conducive to growth. This initial feedback is also characterized by the empathetic impact that emerges, in particular, from the way in which observations on the role of the therapist are formulated, valuing and strengthening their confidence.

In terms of the didactic approach, the AI's output seems flawless: the feedback is formulated in a clear and structured manner, making it easy to read and understand. It offers fairly detailed assessments of the therapeutic techniques used, integrating this assessment with practical alternatives for implementing therapeutic work. A clear example of this is the suggestion to introduce movement- and art-oriented Gestalt experiments, which adds a creative perspective consistent with the patient's needs.

In terms of treatment support, the feedback provides ideas that can be implemented immediately. It suggests ways to address nonverbal resistance and strengthen the patient's resources, while promoting the therapist's professional growth. The ability to identify areas for improvement, such as attention to the balance between support and confrontation, is always expressed in a constructive and encouraging manner. The feedback in question places some emphasis on patient safety and the importance of maintaining a balanced approach that respects boundaries and distances. This emphasis reveals the ethical orientation as a key component of this elaboration. The good adaptation to the therapist's professional level can be seen in the way technical and reflective insights are offered. The feedback in this test made use of the explicit invitation at the end of the prompt to ask further questions, which were asked in a relevant manner; however, this request was not processed when asked at the end of the clinical case presentation. This is the main distinguishing feature between the result of this test and that of the next two. The feedback processed by the ChatGPT4o model used by the ‘expert’ plus account, with LTM function, is characterized by an analytical approach, with greater emphasis on diagnostic and structural analysis. The language produced is clear and respectful to read, but appears less impactful on an empathic level. Communication in this case is more direct and informative. Unlike in the previous test, the feedback processing in this test made appropriate use of invitations to ask further questions, contributing to the structuring of a genuine interaction. The result was a well-organized processing, in which the question-answer structure facilitates understanding and guides the therapist in the analysis of the case. The focus on the analysis of Gestalt techniques is organized around each of the three sessions presented, highlighting their strengths and providing suggestions for possible further developments. Equally appropriately, although more concisely, the aspects of relational dynamics are explored in depth.

While the above suggestions are adequate, the treatment support plan is limited, as it lacks specific suggestions for immediate applicability. The feedback in question, while adequate from an ethical point of view, does not explicitly emphasize its principles.

Finally, a generalist approach emerges that is less suited to the therapist's level of development, neither valuing nor constructively criticizing the skills already demonstrated, but merely illustrating step by step what the therapist has done and what they could do.

The third feedback, generated by the ChatGPT-4o test using a newly created free account without the LTM function, also showed interesting qualities. First of all, like the second and unlike the first, it made good use of the invitation to ask open-ended questions both after the prompt and after the presentation of the case. This fostered a dialogic dynamic that contributed significantly to the improvement in the quality of the feedback. In this case, we are offered a more narrative text in which we find a balance between empathic impact, reflections on theory, and practical advice. It is characterized by language that comes across as welcoming and empathetic, through observations that explicitly value and recognize the therapist's work. The narrative language may require greater attention from the reader than the more direct style that characterizes the previous two feedbacks; however, both the analysis of Gestalt techniques and the practical suggestions are very rich and well articulated. A particular strength of this feedback is the support for treatment: this elaboration offers multiple ideas for the continuation of psychotherapy.

In addition, areas for improvement are highlighted with constructive suggestions that reinforce the sense of collaboration, communicating respect for the pace of the therapist and the patient.

The ethical orientation is articulated in in-depth reflections on patient safety and the therapist's responsibilities. Consideration for the therapist's level of development emerges from the presence of ideas that stimulate continuous development and enhance the skills already demonstrated.

5. Discussion and Conclusion

This study shows that the use of a free and accessible AI model, such as ChatGPT-4o, without LTM, can provide valid and comprehensive results, useful to therapists and trainees with limited resources. Whether free or paid, ChatGPT not only supports but reproduces and sometimes improves the essential functions of clinical supervision [25], challenging the boundaries between human and artificial expertise and raising questions about the nature of expertise and the transmission of knowledge. We have seen how the third feedback stood out for its richness, good articulation, and a sensitive balance between all the central elements of good supervision, such as empathy, theoretical reflections, and practical suggestions. AI integrates different dimensions of supervision—relational, didactic-technical, treatment support, and professional guidance—into a process of epistemological hybridization [26], incorporating complex concepts from Gestalt Therapy. This challenges the boundaries between human experience and algorithmic processing, showing how AI can produce empathetic and relevant feedback.

A crucial role was played by the new prompt developed in this study: by integrating an improvement in the frame instructions provided preliminarily to the chatbot with the elements of good supervisory feedback that emerged from one of the previous studies, we obtained significantly more complete and oriented elaborations than those observed by Cioffi et al. (2024) [19]. This improvement was observed in all three tests carried out in the present study, regardless of the characteristics of the model used. This shows that even free tools can be useful for integrating supervision practices. This evidence represents an important step towards inclusiveness, as free and accessible AI supervision tools can reduce professional hierarchies and democratize expertise, while entailing risks of simplification and standardization of supervision [27]. The use of AI can be seen as cognitive augmentation, extending the therapist's reflective abilities and fostering new forms of conversational supervision. Finally, the study proposes a synthetic epistemology of supervision, recognizing the hybrid nature of clinical knowledge and the possibility of integrating human and artificial skills to create new forms of professional learning.

6. Limitations and Future Developments

One limitation of this study is that the evaluation of the processing was carried out by the researchers who worked on it, including the therapist responsible for the case study, introducing potential observer bias. Furthermore, the main components, with all their characteristics, which were used to construct the new prompt-frame, were extrapolated from a study in which AI-generated supervision feedback (together with a written report by an expert supervisor) was evaluated by an audience of postgraduate students. These limitations lay the groundwork for future developments: the results of a principal component analysis based on the evaluation of supervisory feedback by an audience of more experienced psychotherapists would allow for the creation of a further improved prompt, on the basis of which further tests such as those in this article could be carried out and subjected to a broader and more in-depth evaluation. The findings are based on a single therapeutic approach and one clinical case, significantly limiting generalizability to other psychotherapeutic orientations or diverse clinical presentations. Future research should incorporate rigorous statistical methodologies, including inter-rater reliability assessments and randomized controlled designs comparing AI-assisted with traditional supervision methods.

AI contributes to distributed knowledge ecosystems, where expertise is shared between algorithms, clinical databases, and professional communities [28]. The integration of ethical principles into AI systems highlights the need for new forms of moral and ethical governance, with the participation of developers, professionals, and institutions, as evidenced by studies proposing structured approaches to clinical supervision to mitigate risks and amplify the potential of AI in medical training [29]. The future challenge will be to effectively combine technological innovation and the core values of the healthcare professions.