https://orcid.org/0000-0001-7332-6157
Figures
Abstract
Introduction
Schizophrenia is a complex mental disorder characterized by disturbances in thinking, emotions, and behavior. Emotional dysregulation is a significant clinical issue. Recent years have seen growing interest in the use of conversational agents in mental health care, including for schizophrenia. This study explores the use of a conversational agent as virtual therapist (Terabot) designed to support the recognition and regulation of three emotions (anger, shame, and fear) in hospitalized individuals with schizophrenia. The pilot study aimed to assess the acceptability of this innovative intervention, focusing on protocol implementation, technical performance, and patient experience.
Methods
The study employed a mixed-methods design involving 35 inpatients diagnosed with schizophrenia (ICD-10: F20.0–F20.9). Quantitative evaluation was based on the Brief Psychiatric Rating Scale (BPRS) and a structured acceptability survey, while qualitative thematic analysis was conducted on patients’ open-ended responses and facilitators’ notes. Research procedures followed COREQ standards.
Results
The study involved 35 patients; 34 completed the quantitative scales, and 32 completed the acceptability survey. The quantitative findings suggest that patients generally have a positive attitude toward the analyzed items, particularly Terabot’s clarity of communication, friendly appearance, and the usefulness of its therapeutic exercises, whereas scores for empathy and emotional understanding were only moderate. The qualitative analysis revealed three main areas of patient experience: (1) the therapeutic relationship with Terabot, including both perceived strengths and clear limitations in relational depth, (2) the perceived usefulness of therapeutic exercises, especially relaxation techniques, and (3) Terabot’s physical features, such as its appearance and voice. Patients highlighted contemporary nature of the tool, as well as its limitations, including technical issues (such as freezing and stuttering), rigid communication, insufficient personalization of exercises, and a lack of empathy. Facilitators’ observations confirmed these limitations while also noting patient engagement and openness to technological innovation.
Conclusion
This preliminary study suggests that Terabot may be an acceptable and potentially useful tool for supporting therapeutic work with patients with schizophrenia. However, the findings must be interpreted with caution, given the study’s exploratory design, the time-limited therapeutic intervention, and technical limitations. Enhancements in relational responsiveness, dialogic flexibility, and the range of therapeutic exercises are necessary before drawing more definitive conclusions. Moreover, controlled clinical trials are required to evaluate Terabot’s effectiveness and its capacity to foster a stronger therapeutic alliance.
Citation: Stefaniak I, Gabor-Siatkowska K, Kozłowski M, Janicki A (2026) A conversational agent as a virtual therapist for patients diagnosed with schizophrenia: A preliminary study. PLoS One 21(2): e0343519. https://doi.org/10.1371/journal.pone.0343519
Editor: Domna Banakou, New York University Abu Dhabi, UNITED ARAB EMIRATES
Received: February 10, 2025; Accepted: February 6, 2026; Published: February 27, 2026
Copyright: © 2026 Stefaniak et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All relevant data are within the manuscript and its supporting Information files.
Funding: The author(s) received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Schizophrenia is a chronic mental disorder characterized by the presence of at least two domains (from among delusions, hallucinations, disorganized speech, behavior, and negative symptoms). According to ICD-11, in order for a diagnosis of schizophrenia to be made, the symptoms should persist for a minimum of one month plus allowance for residual phases [1]. Therapy is recommended in the management of patients; cognitive-behavioral therapy for psychosis occupies a special place in guidelines such as those of the WHO and NICE [2]. In recent years, the use of AI in mental healthcare, including psychiatry and psychotherapy, has become an increasingly popular topic [3,4]. Machine learning constitutes a crucial foundation of contemporary AI systems, including those applied in mental healthcare [5]. Such systems offer several advantages that are particularly relevant for the development of therapeutic applications. They can adapt and make self-corrections, thereby better handling unstructured data, which is particularly relevant in therapeutic processes [6,7]. Complex machine learning algorithms have been employed in studies on innovative tools, designed to predict treatment outcomes in individuals with schizophrenia [8] and to develop tools for enhancing psychiatric diagnostics [9]. The application of artificial intelligence in psychotherapy has also been increasingly well described. AI chatbots have been shown to contribute significantly to reducing suicidal, depressive, and anxious thoughts [10,11]. A meta-analysis confirmed the short-term effectiveness of these interventions compared with control conditions, showing statistically significant effects in the following areas: reduction of depressive symptoms, generalized anxiety, specific anxiety symptoms, improved quality of life and well-being, stress reduction, and psychosomatic disorders [12]. A particularly interesting direction of research is the use of artificial intelligence in therapy for people with schizophrenia. Recent developments in technology-assisted interventions for schizophrenia can be broadly classified into three therapeutic types: (1) text-based chatbots, (2) immersive VR-based therapies, and (3) embodied conversational AI agents. Text-driven chatbots are designed to deliver psychological support through writing-based, automated dialogue systems. They have shown preliminary potential for enhancing access to care and support symptom monitoring, motivation, and aspects of cognitive functioning; however, evidence remains early-stage and should be interpreted with caution, particularly in populations with schizophrenia. Examples include the AI-driven mobile intervention PRIME, designed to enhance reward processing and motivation in individuals with recent-onset schizophrenia spectrum disorders, which demonstrated feasibility in a remote randomized controlled trial [13]. Similarly, SchizoBot, a chatbot employing an artificial neural network to deliver cognitive-behavioral strategies, illustrates how conversational systems can be tailored for individuals with schizophrenia [14]. Broader mHealth interventions within this domain address adherence to medication, cognitive training, and symptom management, but the current evidence remains limited by small sample sizes and the scarcity of randomized controlled trials [15]. A scoping review has further highlighted the growing role of chatbots in rehabilitation, including applications in symptom monitoring, pharmacotherapy management, relapse prediction, functional training, and psychosocial support, while also noting the scarcity of randomized controlled trials in this area [16,17]. The second category involves virtual reality (VR)-based interventions, which leverage a high degree of sensory immersion to create environments perceived as real, thereby enabling controlled exposure to scenarios associated with threat or distress. Recent clinical trials have demonstrated that VR-based cognitive therapy, such as the gameChange program, can significantly reduce agoraphobic avoidance, distress in everyday situations, and ideas of reference and persecutory delusions, among individuals with schizophrenia spectrum disorders, especially those with high levels of anxiety and avoidance [18]. VR platforms allow patients to gradually confront feared social environments in a controlled, reproducible manner, enabling therapists to tailor the intensity of exposure and monitor real-time responses. Additionally, VR-based social cognition training programs have shown promising effects on emotion recognition, theory of mind, and functional outcomes, complementing traditional psychosocial interventions. Despite these advances, more research with long-term follow-up and larger samples is required to establish the durability of therapeutic gains and identify moderators of treatment response. Evidence indicates that VR interventions show moderate effectiveness in reducing delusional symptoms and may outperform traditional therapeutic approaches, although further research is needed to clarify underlying mechanisms and establish long-term efficacy [19,20]. The third emerging approach centers on embodied conversational agents, that is, interactive systems equipped with a visual or auditory embodiment designed to engage patients in realistic, relationally meaningful dialogues, an approach exemplified by interventions such as Avatar Therapy [21]. Embodied relational agents, comprises avatar-based and agent-based psychotherapeutic interventions focused on relational and dialogical mechanisms of change. Initiated by Julian Leff, Avatar Therapy enables patients to engage in structured dialogue with a digital representation of their distressing voice, facilitating emotional processing, and a renewed sense of control. Subsequent studies have shown significant reductions in the frequency and distress of auditory hallucinations, improved functioning, and meaningful therapeutic engagement [22–24]. Related approaches, including emerging embodied agents such as Terabot, build on similar principles by offering human-like interactions designed to support emotional expression, relational work, and therapeutic dialogue. From a conceptual standpoint, Terabot aligns most closely with the embodied conversational agent paradigm, as it combines natural-language interaction with an artificial persona intended to support emotional regulation and relational engagement. This places it within the same broader family of technologies that aim not merely to simulate conversation but to recreate aspects of interpersonal exchange relevant to psychotherapeutic processes. The distinction between the three technology-assisted therapeutic approaches used among patients diagnosed with schizophrenia: chatbots, immersive VR-based therapies, and embodied conversational agents, is significant. Each of these tools targets different mechanisms of change and serves distinct therapeutic functions.
The literature also reports several potential adverse effects of artificial intelligence – based interventions among individuals diagnosed with schizophrenia, including misunderstandings of AI-generated content, reinforcement of maladaptive beliefs, and overstimulation resulting from technologically rich environments [25]. These findings underscore the fact that the integration of AI into therapeutic processes for people with psychosis is associated not only with promise but also with a degree of clinical uncertainty. Consequently, the development and implementation of such technologies evoke both hope, regarding improved access, personalization, and scalability of care, and concern about unintended psychological or behavioral consequences. Given this dual landscape, a systematic investigation into the efficacy, safety, and acceptability of these innovative treatment modalities is essential. This includes not only controlled clinical trials assessing therapeutic outcomes but also qualitative research elucidating patient experiences, perceived benefits, and barriers to engagement. Such analyses are vital for identifying the full therapeutic potential of AI-assisted interventions, clarifying their limitations, and informing evidence-based guidelines to ensure their responsible and clinically appropriate use.
It has been observed that despite the substantial increase in research and efforts to evaluate the effectiveness of technologically supported therapeutic methods, there is still an insufficient number of studies examining the use of these methods in groups of patients with severe mental disorders, such as schizophrenia [17].
Emotions in patients with schizophrenia spectrum disorders
Dysfunction in emotional regulation is an important clinical issue in patients with schizophrenia [26]. Gross’s concept of emotions emphasizes that they are dynamic, with regulatory processes playing a key role in their course. Gross distinguishes three essential stages of emotion regulation: situation selection, identification of the need for regulation, and implementation of strategies. Previous research evidence indicates that schizophrenia is associated with abnormalities at each of these stages of emotion regulation [27,28].
Based on Gross’s concept, we analyzed the difficulties experienced by patients with schizophrenia in three stages. It was found that patients with schizophrenia exhibited an ineffective threshold for identifying the need for regulation—they attempted regulation too frequently at low levels of negative affect and too infrequently at high levels, while investing inadequate effort in regulation [29]. It was noted that these patients had a limited repertoire of regulatory strategies and a tendency to choose maladaptive ones [30]. It was further indicated that individuals with schizophrenia demonstrated an impaired implementation of many regulatory strategies, likely resulting from cognitive and neural dysfunctions. This manifested, among other things, in excessive switching between strategies and delayed disengagement compared to control groups [31]. These data became the basis for developing short therapeutic exercises implemented within Terabot. Their design was based on cognitive-behavioral therapy (CBT), one of the most thoroughly researched and widely used methods for the treatment of mental disorders, including schizophrenia [32,33]. Clinical inspiration was also drawn from the so-called “third wave of CBT,” which includes approaches such as Acceptance and Commitment Therapy (ACT), Dialectical Behavior Therapy (DBT), and Mindfulness-Based Cognitive Therapy (MBCT). These approaches emphasize the acceptance of emotional experiences, psychological flexibility, and mindful, conscious responding instead of automatic and maladaptive strategies [34–37]. A direct conceptual foundation was Barlow’s transdiagnostic protocol – the Unified Protocol for Transdiagnostic Treatment of Emotional Disorders – which assumes that many mental disorders, from anxiety and depression to psychotic disorders, share common mechanisms underlying difficulties in emotion regulation [38].
In this study, we assessed the applicability of Terabot virtual therapists in brief interventions aimed at emotion regulation in individuals with schizophrenia.
The primary aim of the study is to evaluate the acceptability of the intervention based on an embodied conversational agent (Terabot) in individuals diagnosed with schizophrenia.
The additional (exploratory) aims are:
- To assess the feasibility of the intervention, including implementability and potential technological barriers.
- To analyze relational dynamics between the participant and the conversational agent, with particular emphasis on the quality of the dialogue.
- To identify potential limitations and risks associated with the use of AI-supported interventions in patients with severe mental disorders.
Methods
Study design
The study was designed as a mixed-methods investigation, combining quantitative and qualitative approaches to obtain a more comprehensive picture of patients’ experiences in therapy with Terabot. The qualitative appraisal of the study was conducted with reference to the COREQ criteria (Consolidated Criteria for Reporting Qualitative Research) [39], with a full report provided in S1 Table. The reporting standard was applied consistently, both in the preparatory phase (focus group and development of the therapeutic protocol) and in the active research phase (implementation of the intervention).
To characterize the study group, descriptive statistical analysis was conducted, enabling a concise and transparent presentation of key sociodemographic and clinical features. Statistical analyses were conducted using IBM SPSS Statistics 30. Basic descriptive statistics, along with the Shapiro–Wilk test and the Student’s t-test for dependent samples, were performed. The level of significance in this chapter was set at α = 0.05. In parallel, a qualitative thematic analysis was performed on patients’ survey responses and facilitators’ observational notes, allowing an in-depth exploration of participants’ experiences and perceptions of the intervention.
Criteria for inclusion
The study was approved on 27 April 2022 by the Ethics Committee of the Institute of Psychology and Neurology in Warsaw, Poland; resolution No. IV/2022, and it conforms to the provisions of the Declaration of Helsinki. The research was conducted between April 2022 and August 2023. Participants were recruited from the Psychosis Relapse Prevention Ward at the Warsaw Institute of Psychiatry and Neurology. The ward admits inpatients with severe mental disorders who are stabilized and in partial or complete remission, with the primary objective of preventing further mental health crises.
The ward offers a comprehensive therapeutic program that includes community meetings, psychoeducation sessions, therapy groups, metacognitive training, initiative workshops, physical activities, and cognitive training. Within this therapeutic environment, patients who additionally met the study’s inclusion criteria were invited to participate in supplementary sessions with the virtual therapist (Terabot).
Eligibility for the Terabot intervention was restricted to patients with diagnoses falling under ICD-10 codes F20.0–F20.9, confirmed by experienced psychiatrists prior to hospital admission [40]. The detailed inclusion and exclusion criteria are provided in Table 1. All participants received thorough information regarding the aim, duration, participation requirements, and potential benefits of the study, accompanied by a full confidentiality clause. Personal data were anonymized.
Preparatory phase
Interest in the issue of emotion regulation in patients with schizophrenia dates back to the 1990s, while the last two decades have seen intensive development of social skills and cognitive training, of which emotion regulation was an important component. Increasingly, studies indicate that improving emotion regulation can contribute to a reduction in negative symptoms and improved social functioning [41]. In the rehabilitation ward where the study was conducted, emotion regulation training appeared as part of broader therapeutic programs; there had been no separate module focused solely on this issue [42]. These observations became the inspiration for developing Terabot – a tool designed to support patients in the independent training of emotional regulation [43]. The therapeutic program was based on James Gross’s concept and included elements of psychoeducation, emotion recognition, regulatory exercises (e.g., relaxation and breathing techniques), reframing emotional meanings, and individual tasks to consolidate new ways of coping with emotional experiences.
Therapeutic protocol.
During the preparation phase, a focus group was conducted to identify the emotions that would subsequently be targeted in therapeutic work. The meeting of the group took place in a psychiatric ward among patients diagnosed with schizophrenia (according to ICD-10: F20.0–F20.9), in the same ward where the study was conducted. It was a one-time meeting led by two clinicians. One of the facilitators of the session was a clinician (this individual had extensive clinical experience as a psychiatrist, psychotherapist, and supervisor), and the other person (with a background in psychology) had an assisting role. The meeting lasted two hours. The first step involved providing a short discussion on emotions. They were subsequently asked to discuss the emotions they experienced most frequently and the strategies they used to cope with them. Patients’ statements were recorded by the assisting clinician, who did not influence their responses. The therapeutic team then evaluated the collected material to determine which emotions posed the greatest challenges for patients. A thematic analysis of the notes was carried out, leading to the selection of three emotions for the development of therapeutic dialogues and exercises. The group consisted of 10 inpatients.
Design and development of a polish-language therapeutic dialogue system.
An initial motivation for creating Terabot was the need to reduce the workload of therapeutic staff in rehabilitation wards, where professionals face significant demands. Terabot was intended to serve as an independent tool that patients could use at any time during their stay, allowing for the systematic development of emotion regulation skills. The first phase included creating a virtual therapist to conduct independent patient conversations. It was a spoken dialogue system for the Polish language based on the RASA platform [44]. The system has been initiated with sample dialogues and then trained during the so-called conversation-driven development. A simultaneous process was the development of the therapeutic protocol. Once the preliminary version of Terabot and the therapeutic protocol were developed, the next step involved conducting a preliminary examination test with a non-clinical group. This stage aimed to improve Terabot’s communication skills and teach him to react to as many responses as possible. After completing this stage, the clinical phase was initiated.
Description of the role of facilitators during the sessions with Terabot.
Two facilitators were trained in their tasks for the study. Neither of them had clinical experience. One of them was always present during the entirety of each meeting, supporting the patient in difficulties that might arise during sessions with Terabot. The facilitator’s tasks included leading the patient into the room, assisting during subsequent parts of the session (e.g., prompting the patient to speak more loudly or clearly), and systematically observing and documenting any technical difficulties, emotional reactions, or potential adverse events arising during the meetings with Terabot, including signs of distress, symptom exacerbation, or behavioral changes. Facilitators also explained the content of the individual assignments to the patients. They did not know the participants beforehand and ensured that their responses remained unbiased. The participants were informed about the purpose of the study and their right to withdraw at any time. Each participant consented to the presence of the facilitator. The facilitator took notes on the course of the sessions and described the difficulties that arose during the meetings. Importantly, the facilitator did not provide any therapeutic, emotional, or clinical support and did not intervene in the therapeutic process; their role was strictly limited to technical, organizational, and observational tasks.
Terabot as a virtual therapist
Terabot was a task-oriented spoken dialogue system for the Polish language, developed in Poland at the Warsaw University of Technology [45–47]. By offering therapeutic spoken interaction in the Polish language, Terabot contributes to filling a gap in existing dialogue-system research. At the hospital, the therapeutic station consisted of a computer with a screen, a microphone, and speakers. A photograph of this station is shown (Fig 1).
The mouse was used only to turn on the operating system and the dialogue system. For conducting the conversations, it was hidden on the left side (where the assistant was sitting), being invisible to the patient. The system was equipped with a speech synthesizer and speech recognition system and was therefore capable of communicating with the patients verbally. The architecture of Terabot is shown (Fig 2).
Terabot recognized the patient’s utterances (converted speech into text), interpreted them using the dialogue system engine, and returned an answer in the form of a text compliant with the therapy exercise scenarios provided earlier. A speech synthesizer then converted the answer into speech and played it through the speakers. A computer dialogue system automatically processed the patients’ statements in the following manner:
- intent analysis: the system recognized the intentions behind the patient’s utterance (e.g., recounting an event, sharing emotions, confirming/negating a statement) using the DIET Classifier algorithm [48],
- keyword analysis: detecting and identifying the type of exercise to be performed.
Based on its content, this dialogue system automatically recognized the tone (sentiment) of the patient’s statement (neutral; positive, e.g., happiness; negative, e.g., anger, anxiety); this is done by an algorithm employing an artificial neural network [49,50].
Physical features of Terabot
While designing Terabot, efforts were made to ensure its appearance was neutral to avoid triggering patients’ biases [51], the appearance is shown (Fig 3). He moved slightly, and his eyes blinked naturally. This gave the patient the impression of sitting in front of the computer during an online session with a therapist. We present its main physical characteristics in Table 2.
Translation from the Polish language: Say “hello” to start a conversation with Terabot.
Description of the intervention using Terabot
Terabot proposed exercises aimed at conducting a conversation about emotions, increasing awareness and improving the ability to recognize them, and then performing an exercise designed to reduce the intensity of emotion. All the commands and tasks were guided by Terabot. The whole intervention was planned to last five days. During this time the patient conversed with the dialogue agent and participated exercises to improve their ability to deal with emotions. The meetings lasted approximately 20–30 minutes. After completing the exercise, the patient was given an assignment related to the issue discussed in the meeting (personal task).
This dialogue system was prepared (taught) to conduct three therapeutic exercises focused on management of anger, anxiety and shame. The emotions included in the therapy protocols were selected based on a demand analysis conducted during a focus group. The exercises were developed on the basis of principles of cognitive-behavioral therapy and mindfulness-based therapy (Mindfulness Based Cognitive Therapy, MBCT) [52].
Exercises for each of these emotions are built on the same principles:
- Greeting.
- Terabot asks if the patient agrees to participate in the exercise.
- Introduction – specifying which emotion the conversation will refer to.
- Asking the patient to recall a situation in which they felt the specific emotion.
- Psychoeducation on the topic of emotion, including its physiological aspects and impact on behavior.
- Proposing a short exercise aimed at helping the patient cope with a particular emotion.
- Personal task assignment.
Terabot was also able to handle cases out of the main scenario. If the patient avoided answering the question, e.g., “Do you recall a situation where you felt angry?”, the system offered pre-defined examples, such as “Imagine you come home hungry and realize you forgot to do the shopping.” The patient was allowed to quit the conversation at any point by expressing such a wish. In such a case, the system suggested meeting at another time and said goodbye. Terabot was also able to answer basic questions about the system itself, such as “Who are you?” “Which languages do you speak?”, “Which topics can you discuss?” Following the psychoeducation part on the topic of emotion, the system offered patients a relaxation exercise. It consisted of several relaxation instructions, interleaved with silence. Afterwards, Terabot asked the patients how they felt. If the patients felt well, Terabot thanked them and said goodbye. If they answered they had problems with relaxation, Terabot asked for details and encouraged them to join another session with Terabot.
Table 3 presents the structure of the Terabot intervention, psychotherapeutic stages, and its corresponding stages of emotion regulation according to Gross’s model. The individual dialogue components operationalize, in sequence: the following antecedent-focused strategies (identifying the emotion and its triggering situation), attentional deployment and cognitive reappraisal (psychoeducation, analysis of physiological reactions, assigning new meaning to emotions), as well as response modulation through relaxation and mindfulness exercises. The intervention concludes with an evaluation of the effectiveness of the emotion regulation process and a homework assignment aimed at consolidating and generalizing the acquired skills.
Scales and survey used in the study
Sociodemographic data were collected using a questionnaire prepared by the research team, which included the following information: age, gender, education, marital status, employment status, and place of residence. Treatment-related information collected included diagnosis, number of hospitalizations, and current treatment. An acceptability survey was prepared to evaluate the use of a conversational agent, which patients completed after their interactions with Terabot. The Davis technique was used to assess the validity of the items developed [53]. The questions were assessed by four independent experts, comprising two clinicians and two technology specialists. Each question was rated on a scale from 1 to 3 (1 – not relevant, 2 – useful but not essential, 3 – essential). Calculations were performed for each question, and only items with a content validity index > 0.80 were included in the final survey. The internal consistency of the acceptability survey (Q1-Q12) in the present sample was Cronbach’s α = 0.84. Below are the questions used to measure the acceptability of the use of the dialogue agent:
- Q1. The statements made by Terabot were understandable to me.
- Q2. I felt that Terabot understood what I wanted to say.
- Q3. I felt that Terabot understood my emotions.
- Q4. The questions asked by Terabot were appropriate.
- Q5. The exercises proposed by Terabot were helpful.
- Q6. The statements made by Terabot sounded natural.
- Q7. Terabot spoke too fast for me.
- Q8. I enjoy talking with Terabot.
- Q9. The presence of the observer was helpful to me.
- Q10. Terabot looks friendly.
- Q11. I feel that therapy with Terabot helped me.
- Q12. Computer-based methods, such as Terabot, can be used in therapy.
Patients related to the above statements on a 5-point Likert Scale (1 - totally disagree, 2 - disagree, 3 - hard to say, 4 - agree, 5 - totally agree).
At the end of the survey, we included open-ended questions:
- What did you like the most about the exercise with Terabot?
- What did you like the least about the exercise with Terabot?
- If you wish, you can add your additional comment.
The overall severity of psychopathological symptoms was measured with the standard version of the BPRS (Brief Psychiatric Rating Scale) [54]. The BPRS is one of the most widely used scales for the measurement of psychopathologies. It includes 18 symptoms assessed by a clinician on a 7-point scale (1 = symptom not present; 7 = symptom extremely severe) (Cronbach’s α = 0.72). An additional element of the qualitative assessment consisted of notes taken by the facilitator and presented during each session in which the patient interacted with Terabot. There were three types of issues observed by the facilitators:
- Technical Issues and Functionality
- Quality and Personalization of Interaction
- Facilitator’s presence during sessions with Terabot.
Safety monitoring and adverse events
Adverse events were operationally defined as any unintended deterioration in participants’ mental state, including: (1) clinically significant exacerbation of psychotic symptoms, (2) emergence or disclosure of suicidal ideation and an associated increase in suicide risk, and (3) heightened levels of anxiety and psychomotor agitation. Safety monitoring was conducted indirectly through the administration of the Brief Psychiatric Rating Scale (BPRS) at baseline and following completion of the intervention. Additionally, qualitative safety monitoring was supported by systematic observations recorded by a facilitator present during each session. The facilitator documented participants’ behavioral and emotional responses throughout the meetings. In cases where potential adverse events were identified, the facilitator was authorized to notify the patient’s treating psychiatrist regarding any observed deterioration in mental state, thereby enabling appropriate clinical follow-up.
Results
Socio-demographic data
The study group comprised 35 patients (20 men and 15 women). The mean age of respondents was 37.1 years (SD = 11.1), ranging from 22 to 64 years. The average number of hospitalizations was 6.1 (SD = 5.0), with values ranging from 1 to 20. All patients had F20.0 - F20.9 diagnoses according to ICD-10 (according to ICD-10: F20.0). All patients were being treated with antipsychotic medications. Table 4 below shows descriptive statistics for age and number of hospitalizations.
Table 5 below presents the basic sociodemographic data of the respondents. The majority of the participants were male (57.1%) and single (77.1%). More than half were retired or on disability pension (51.4%). The largest proportion of participants (31.4%) reported residing in cities with populations exceeding 100,000. The majority of patients had either secondary (51.4%) or higher (31.4%) education.
BPRS
This scale was used to monitor and confirm the stability of participants’ mental state during the study. BPRS data were available for 34 participants. One patient did not complete the scale and was therefore excluded from the analysis. The Shapiro–Wilk test did not reveal significant deviations from the normal distribution. The analysis revealed no statistically significant differences in the BPRS scales (BPRS T0: M=2.15; SD=0.51; BPRS T1: M=2.02; SD=0.57; t=1.44; df=33; p=0.16).
Analysis of the acceptability survey and experiences with Terabot therapy – quantitative and qualitative analyses
The assessment was performed by analyzing the responses to closed-ended survey questions, analyzing the open questions in the survey, and the facilitator’s observation. Fig 4 and Table 6 present the patients’ responses to closed–ended questions. We obtained responses from 32 patients participating in the study. Three people refused to complete the survey.
M – mean, SD – standard deviation, Q1-Q12 – survey questions (1-12) posed to patients.
Most questions received neutral and positive ratings – the responses “hard to say” (3) and “agree” (4) were the most dominant. The mean values ranged mainly between 3.1 and 4.3, indicating that patients were generally positively inclined, although not unanimous in all aspects. The highest-rated questions (mean approx. 4.2–4.3) show that participants were satisfied with the therapy. Questions with lower averages (approx. 3.0–3.1) suggest that opinions in these areas were divided – some patients were neutral or dissatisfied. The analysis included one reverse-coded question: “Terabot spoke too fast for me.” In this case, a higher score indicated greater agreement with this statement. Standard deviations of 1 or higher indicate considerable variability in opinions on certain questions.
Qualitative analysis was exploratory and based on thematic content analysis of open-ended survey responses and facilitators’ notes. Two researchers independently coded the data, developed categories and themes, and resolved discrepancies through discussion. Researcher triangulation was applied to enhance credibility, and illustrative participant quotations were included in the results. The analysts acknowledged the potential influence of their clinical and research backgrounds, including the fact that one study leader was employed in the ward where the research was conducted. Despite the pilot nature of the study and the limited dataset, recurring patterns were identified, indicating preliminary thematic convergence. The analysis was conducted manually without the use of specialized software. Three overarching themes emerged: (1) physical characteristics of Terabot, (2) relational and communicative aspects of the interaction, and (3) perceived therapeutic usefulness of the exercises [55].
Based on the responses to the open-ended questions (“What did you like most about the exercise with Terabot?”), the following issues have been identified in several areas, see Table 7. We analyzed the responses in three categories. Table 7 shows examples of patients’ statements from each theme and category.
In response to the question about what patients liked most in working with Terabot, they most often emphasized the relationship and communication with the virtual therapist (e.g., the possibility of conversation, being understood, and its calm manner of speaking). The second important aspect was the therapeutic program itself, namely the proposed exercises and relaxation methods. Patients also drew attention to Terabot’s physical features – including its appearance, voice, and way of presenting itself. Among the responses classified as “other”, the most frequently mentioned was the contemporary character of the form of therapy and the opportunity to try something new. When conducting the thematic analysis of responses to the question about what patients liked the least in working with Terabot, we also aimed to distinguish the three categories described. See the following Table 8.
The qualitative findings indicate that patients’ experiences with Terabot were shaped above all by the perceived therapeutic relationship. Interestingly, this aspect was mentioned both as the most and the least satisfying element, suggesting that the quality of relational experience with the virtual therapist was central to a patient’s evaluation. The therapeutic exercises and relaxation methods were also valued, though for some participants they represented a source of dissatisfaction. Physical features of Terabot, while less prominent, contributed in both positive and negative ways. Taken together, the results highlight that even in technology-mediated interventions, the relational dimension remains crucial, while content and form play a secondary yet meaningful role. In the survey, patients could also provide their own free comments about the therapy with Terabot, as presented in Table 9 below.
Qualitative analysis – facilitators’ observations
In the next step, the facilitator’s comments were subjected to thematic analysis. Three main thematic areas were distinguished: (a) technical issues and functionality, (b) quality and personalization of interaction, and (c) the role of the facilitator’s presence during sessions:
- a). Technical Issues and Functionality:
- Terabot faced difficulties understanding certain responses, particularly lengthy, unclear, or stereotypical ones.There were issues with low reactiveness, including unrelated or non-empathic responses, lack of emotional engagement, and failure to acknowledge patient input.
- Errors during tasks necessitated restarts, indicating a lack of task robustness.
- Terabot struggled to guide patients through recalling emotional memories or facilitating deeper discussions.
- b). Quality and Personalization of Interaction
- Behavioral and communication barriers:
- Patients exhibited passive or non-engaged behavior, such as failing to complete tasks, providing short or irrelevant responses, or outright ignoring Terabot.
- Some patients expressed discomfort or frustration, including speaking quietly, shouting, or showing delayed responses.
- Cognitive challenges, such as struggling to recall emotional memories or maintaining focus during exercises, impacted interactions.
- Emotional and Relational Disconnect:
- Patients expressed embarrassment or discomfort in communicating with a machine, especially one they perceived as “non-human” or lacking empathy.
- Some patients expected an “in-depth conversation” but were disappointed by Terabot’s limitations.
- Patients displayed avoidance behaviors (e.g., looking down, truculence) and skepticism about Terabot’s role.
- c). Facilitator’s presence during sessions with Terabot:
- For most patients, the facilitator’s presence was not bothersome, suggesting that it did not interfere with the interaction.
- In some cases, however, patients turned to the facilitator for reassurance, reflecting a need for human validation and encouragement.
The analysis of patients’ open-ended responses and the facilitator’s observations indicates that the experiences related to using Terabot were diverse and included both positive elements and limitations in the functionality of this tool. In open-ended responses, patients pointed out its physical features as crucial during the exercises. Some patients confirmed that the physical features of Terabot (e.g., natural manner, voice, movement) enhance patients’ positive experiences and increase their engagement in therapy. Some patients give positive evaluations of the innovation and uniqueness of this type of therapy. The therapeutic exercises were perceived as effective and practical, with a strong focus on relaxation and their potential application in everyday situations. One of the most essential elements that the patients mentioned was the therapeutic relationship with Terabot. Among the frequently mentioned aspects were the fluency of this relationship and its deepening. These elements were deemed significant both as aspects appreciated by patients and as areas identified for improvement. Technical issues, such as freezing and stuttering, were significant limitations to Terabot’s functionality. Patients also noticed artificiality in its behavior and visual aspects. Patients experienced a lack of fluency in communication, limitations in understanding their responses, and inflexibility in dialogue, which negatively impacted their relationship with Terabot. Some patients found the exercises insufficiently tailored and lacking variety. Limiting interactions to exercises alone was perceived as a lack of engagement and flexibility.
In this final section, we have analyzed the distinct areas in the opinions of patients and the facilitators, as well as those that overlapped. Both patients and the facilitator recognized Terabot’s limitations in communication, empathy, and task execution. Both groups agreed that the tool requires improvements to better meet user needs. Patients placed greater emphasis on Terabot’s potential and highlighted the positive aspects of its operation, while the facilitator focused on a critical analysis of its shortcomings. The patients’ response to the presence of the facilitator was not fully reflected in their feedback.
Safety monitoring and adverse events
The analysis revealed no statistically significant differences between pre- and post-intervention BPRS scores, suggesting the absence of overall deterioration in participants’ mental state over the short observation period. Qualitative data and conversation transcripts did not indicate the presence of suicidal ideation or a worsening of psychotic symptoms. Facilitator observations showed that some patients experienced episodes of dysphoria and frustration (e.g., raised voice); however, these reactions did not lead to sustained mental health deterioration or necessitate session termination by the facilitator. Consequently, these events were not formally classified as adverse events, which limits the possibility of their quantitative assessment. Overall, the collected data provide preliminary evidence supporting the short-term safety of the intervention; nevertheless, the lack of systematic adverse event monitoring and formal escalation procedures warrants cautious interpretation of the findings.
Discussion
The present study examined the acceptability of a dialogue-based agent (Terabot) used in brief emotion-focused interventions for patients diagnosed with schizophrenia. Overall, the findings suggest that Terabot was perceived as an acceptable intervention; however, patients also identified several limitations that may constrain its therapeutic usefulness and inform directions for future development. Patient satisfaction emerged as one of the key indicators of Terabot’s acceptability, as reflected in the quantitative results of the acceptability survey. As shown in previous research, patients tend to perceive technology-supported therapeutic interventions, such as those delivered by embodied conversational agents, as modern and attractive forms of treatment [56].
Integrating qualitative and quantitative findings, it becomes evident that communication, the physical features of Terabot, and the nature of the therapeutic exercises played a central role in shaping patients’ experiences. These aspects were consistently highlighted in open-ended responses addressing both positive and negative experiences of interacting with Terabot, and similar observations were reported by the facilitators. Notably, the perceived therapeutic relationship with the virtual therapist appeared to be the most influential dimension. Relational and communicative elements -such as the opportunity for dialogue, a sense of being understood, and Terabot’s calm manner of speaking – were frequently valued by participants.
One of the key findings is the importance of Terabot’s physical features, such as its appearance and the timbre of its voice; the vast majority of patients rated Terabot’s appearance as friendly and its voice as soothing. These results are consistent with studies on the impact of “appearance” and “voice quality” in human – computer interaction, indicating that a neutral and natural appearance combined with an appropriately selected voice can foster a sense of safety and enhance patients’ acceptance of the tool [57].
The uncanny valley phenomenon provides a useful interpretative framework for analyzing patients’ interactions with Terabot [58,59]. Terabot’s visual representation was based on a video recording of a human actor with a partially covered face, a design choice intended to convey neutrality, safety, and professionalism. However, the combination of human facial features with limited facial expressiveness (due to the face mask) may have generated perceptual ambiguity in the studied patient group, making it more difficult to infer the agent’s intentions. Given Terabot’s therapeutic objectives of supporting the emotion regulation process, the limited legibility of the virtual therapist’s emotional cues and communicative intent may have attenuated the clarity of the virtual therapist’s emotional states and intentions, thereby attenuating the expected effects of the intervention [60]. From the perspective of Bordin’s model of the therapeutic alliance, reduced clarity in the agent’s affective signaling may have selectively weakened the emotional bond component of the alliance, despite participants’ generally positive evaluations of the intervention’s structure and of the therapeutic goals and tasks it clearly articulated [61]. Empirical research on the uncanny valley further suggests that experiential discomfort and potential clinical risk are more likely to arise when users construe conversational systems as autonomous relational agents endowed with empathic and mental capacities, while these systems lack the flexibility and affective attunement characteristic of human interaction. Within this framework, attempts to enhance Terabot’s anthropomorphic qualities, although conceptually aligned with efforts to foster relational engagement, may have inadvertently elevated users’ expectations beyond the agent’s actual dialogical and regulatory capacities, resulting in perceived artificiality or emotional distancing in a subset of participants [62].
Although the present study did not employ a dedicated measure of the uncanny valley effect, several of its constitutive elements, including perceived naturalness, affective attunement, and interpersonal comfort, were reflected in both quantitative acceptability indices and qualitative user reports. Collectively, this effectiveness may depend less on maximizing human likeness per se than on ensuring coherence between the agent’s embodied form, communicative behavior, and functional dialogical capabilities, particularly when deployed in clinical populations diagnosed with schizophrenia.
The quantitative findings suggest that Terabot’s main strengths lie in the clarity of its communication, its friendly appearance, an appropriate speaking pace, and the perceived usefulness of some of the therapeutic exercises. Although mentioned less frequently, Terabot’s physical characteristics, such as its appearance, voice, and mode of presentation, also contributed to both positive and negative perceptions. The therapeutic content, including exercises and relaxation techniques, was generally regarded as useful and applicable; however, some participants perceived these elements as insufficiently tailored or lacking variety. These findings are consistent with earlier studies on short digital interventions, where excessive schematization was found to reduce engagement and therapeutic effectiveness [58]. The literature emphasizes that the effectiveness of computer-based conversational agents largely depends on their ability to build relationships with users. In the article [57] the authors define relational agents as computational systems designed to establish and maintain long-term socio-emotional relationships, grounded in mechanisms known from social psychology and interpersonal communication. The authors demonstrate that relational behaviors exhibited by agents, such as empathy, appropriate communication strategies, or the continuation of themes from previous interactions, enhance trust, engagement, and the willingness to continue the interaction. This perspective is particularly relevant in therapeutic interventions, where the quality of the relationship (working alliance) constitutes one of the key predictors of therapeutic outcomes. The significance of the therapeutic relationship in interactions with conversational agents has also been confirmed in prior research on the working alliance in digital therapy. It has been demonstrated that the quality of the relationship with a virtual agent directly influences patients’ engagement, satisfaction with therapy, and adherence to recommendations [57]. Similar conclusions were drawn in a scoping review of embodied conversational agents in clinical psychology [56]. More broadly, it has been argued that the therapeutic relationship remains one of the most powerful predictors of treatment outcomes, regardless of modality, and is equally relevant in digital contexts. In the case of schizophrenia, these findings are particularly meaningful, as impairments in interpersonal relationships are among the core difficulties faced by this patient group [59]. Thus, the possibility of establishing a satisfactory and “safe” therapeutic relationship with a virtual therapist may serve not only as an element supporting therapy but also as a therapeutic goal in itself.
Empathy is closely linked to the capacity for affective attunement, that is, the ability to respond precisely to the emotional tone of the patient’s communication [63]. The rigid, formulaic responses generated by Terabot make such attunement difficult, potentially weakening the patient’s sense of being genuinely understood [64]. These communicative limitations may also lead to disruptions in the therapeutic alliance, as a lack of flexibility and responsiveness is often interpreted as a sign of poor fit, disengagement, or insufficient interest [65]. In the context of relational engagement, it is useful to refer to Bordin’s model, which conceptualizes the therapeutic alliance as consisting of three components: agreement on goals, agreement on tasks, and the emotional bond [61]. Contemporary conversational agent research highlights the point that the design of social cues -including qualities of voice qualities, facial expression, embodied behavior, timing, and discourse style – plays a critical role in shaping the user’s perception of rapport and the alliance. Social cues structure how users interpret an agent’s social presence and relational intent; richer cue design is associated with increased feelings of affiliation, trust, and engagement [66,67,68]. This implies that an alliance is not solely an emergent property of the content of dialogue but is co-determined by the designed of human factors: appearance, vocal attributes, responsiveness, and conversational latency.
Terabot performs well in the first two areas due to the clear structure and transparency of its intervention; however, limited emotional attunement and rigid interaction patterns may weaken the relational component, which is essential for fostering safety and trust. Moreover, the absence of follow-up questions, which in natural conversation serve as mechanisms of mirroring and implicit validation, results in a reduced number of signals that typically communicate understanding and acceptance of the patient’s experiences. As a consequence, the interaction becomes rigid, limiting the depth of emotional exploration.
Terabot belongs to the broader class of relational agents but applies relational dynamics within a therapeutic domain distinct from avatar-based and immersive interventions. Whereas avatar therapy is designed for individuals experiencing persistent auditory hallucinations and focuses on externalizing voices and enhancing perceived control, Terabot targets emotion-regulation processes and aims to modify maladaptive coping strategies, positioning it as a more transdiagnostic tool applicable across psychiatric conditions. Unlike clinician-guided avatar therapy or virtual reality–based approaches that rely primarily on sensory immersion and exposure mechanisms, Terabot operates as an autonomous conversational agent capable of sustaining dialogue independently [69–71]. Despite the absence of a human therapist and immersive technologies, patients in the present study highlighted the importance of relational and communicative aspects of interaction, suggesting that meaningful therapeutic engagement can arise through conversational and relational processes alone. Terabot operates through a fundamentally different therapeutic mechanism, grounded in conversational interaction with an embodied conversational agent. The present findings indicate that relational and dialogic processes may play a central role in shaping user experience and acceptability, even in the absence of immersive environments.
shame).
Taken together, these findings suggest that therapeutic conversational agents should be conceptualized not merely as delivery mechanisms for structured interventions, but as relational systems whose alliance-forming capacities depend on the quality, consistency, and multimodality of their social cues. Future work should thus integrate alliance theory from psychotherapy with design frameworks from human – computer interaction to more systematically examine how users form, maintain, or rupture relational bonds with autonomous agents.
Limitations of the study
One of the key limitations of the study concerns the absence of an objective measure of emotion regulation (ER), which prevents any reliable assessment of change in the targeted emotional processes. Without standardized ER scales, the evaluation relied primarily on subjective self-reports, which introduces the risk of inflated self-evaluation, particularly in interactions involving socially desirable responding or perceived expectations from the system. This methodological constraint limits the interpretability of the observed effects and restricts the ability to draw conclusions about the intervention’s actual impact on emotion regulation capacities. Consequently, the generalizability of the findings is reduced, and the results should be viewed as preliminary indicators of acceptability rather than evidence of therapeutic efficacy. Future studies should therefore incorporate validated emotion regulation measures and multimodal assessment strategies to allow for more robust inferences regarding clinical change. Another limitation concerns the relatively small sample size, which restricts the generalizability of the findings. Furthermore, the study was of a preliminary and exploratory nature, which means that the results should be interpreted with caution. Technical issues, such as system freezes or limited dialogue flexibility, may have led to bias in patient experiences and therefore affected the outcomes of the evaluation. Finally, the short intervention period did not allow for the observation of longer-term effects or sustainability of potential therapeutic benefits.
Possible future research
The study presented here is one of the first to assess interventions conducted by a virtual therapist in a group of patients diagnosed with schizophrenia spectrum disorders, evaluating a brief intervention aimed at helping patients manage their emotions. The uncanny valley phenomenon may represent an important direction for future research on therapeutic conversational agents in populations diagnosed with schizophrenia. Systematic investigation of how varying levels of conversational agent realism, emotional expressiveness, and perceived autonomy interact with individuals experiencing psychotic disorders is warranted. Research designed in this manner may contribute to the development of therapeutic tools that effectively support clinical interventions while being optimized in terms of safety, user engagement, and the quality of the therapeutic alliance. Future studies should therefore seek to address questions concerning the optimal degree of anthropomorphism and realism, and their respective effects on acceptability, emotion regulation, and the formation of the therapeutic alliance in clinical populations.
Future research should focus on refining both the technical and therapeutic components of the intervention, with particular emphasis on improving Terabot’s communicative flexibility and empathic responsiveness. The qualitative findings highlight the need for greater conversational scaffolding, including guiding questions and prompts that support more elaborated emotional narratives. Incorporating brief preparatory elements, such as emotion recognition tasks, as well as follow-up questions, may facilitate deeper engagement and more meaningful therapeutic interaction. In subsequent steps, a randomized clinical trial would be necessary. Patients appreciate the modernity and potential of Terabot, but point out specific areas for improvement, such as understanding, the personalization of exercises, and the naturalness of interactions. Further development of the project should take these needs into account to increase its effectiveness and acceptance among users.
Supporting information
S2. Table. This is the S2 Table Terabot Ex. Survey.
https://doi.org/10.1371/journal.pone.0343519.s002
(XLSX)
Acknowledgments
We would like to thank the medical staff from the First Department of Psychiatry of the Institute of Psychiatry and Neurology in Warsaw for their contribution to the clinical assessment of the study participants.
References
- 1. World Health Organization. International statistical classification of diseases and related health problems, 11th revision (ICD-11). https://icd.who.int/browse11
- 2. National Institute for Health and Care Excellence. Psychosis and schizophrenia in adults: prevention and management. 2014. https://www.nice.org.uk/guidance/cg178
- 3. Spytska L. The use of artificial intelligence in psychotherapy: development of intelligent therapeutic systems. BMC Psychol. 2025;13(1):175. pmid:40022267
- 4. Zhou S, Zhao J, Zhang L. Application of artificial intelligence on psychological interventions and diagnosis: an overview. Front Psychiatry. 2022;13:811665.
- 5. Kalanderian HNHA. Artificial intelligence in psychiatry. Curr Psychiatr. 2019;18(8):33–8.
- 6. Jordan MI, Mitchell TM. Machine learning: Trends, perspectives, and prospects. Science. 2015;349(6245):255–60. pmid:26185243
- 7. Li H, Zhang R, Lee Y-C, Kraut RE, Mohr DC. Systematic review and meta-analysis of AI-based conversational agents for promoting mental health and well-being. NPJ Digit Med. 2023;6(1):236. pmid:38114588
- 8. Del Fabro L, Bondi E, Serio F, Maggioni E, D’Agostino A, Brambilla P. Machine learning methods to predict outcomes of pharmacological treatment in psychosis. Transl Psychiatry. 2023;13(1):75. pmid:36864017
- 9. Chen ZS, Kulkarni PP, Galatzer-Levy IR, Bigio B, Nasca C, Zhang Y. Modern views of machine learning for precision psychiatry. Patterns (N Y). 2022;3(11):100602. pmid:36419447
- 10. Dimeff LA, Jobes DA, Koerner K, Kako N, Jerome T, Kelley-Brimer A, et al. Using a Tablet-Based App to Deliver Evidence-Based Practices for Suicidal Patients in the Emergency Department: Pilot Randomized Controlled Trial. JMIR Ment Health. 2021;8(3):e23022. pmid:33646129
- 11. Fulmer R, Joerin A, Gentile B, Lakerink L, Rauws M. Using Psychological Artificial Intelligence (Tess) to Relieve Symptoms of Depression and Anxiety: Randomized Controlled Trial. JMIR Ment Health. 2018;5(4):e64. pmid:30545815
- 12. He Y, Yang L, Qian C, Li T, Su Z, Zhang Q, et al. Conversational Agent Interventions for Mental Health Problems: Systematic Review and Meta-analysis of Randomized Controlled Trials. J Med Internet Res. 2023;25:e43862. pmid:37115595
- 13. Schlosser DA, Campellone TR, Truong B, Etter K, Vergani S, Komaiko K, et al. Efficacy of PRIME, a Mobile App Intervention Designed to Improve Motivation in Young People With Schizophrenia. Schizophr Bull. 2018;44(5):1010–20. pmid:29939367
- 14. Nwoye EO, Muslehat AA, Umeh C, Okodeh SO, Woo WL. SchizoBot: Delivering Cognitive Behavioural Therapy for Augmented Management of Schizophrenia. Digital Technologies Research and Applications. 2024;3(2):1–17.
- 15. Gire N, Farooq S, Naeem F, Duxbury J, McKeown M, Kundi PS, et al. mHealth based interventions for the assessment and treatment of psychotic disorders: a systematic review. Mhealth. 2017;3:33. pmid:28894743
- 16. Boucher EM, Harake NR, Ward HE, Stoeckl SE, Vargas J, Minkel J, et al. Artificially intelligent chatbots in digital mental health interventions: a review. Expert Rev Med Devices. 2021;18(sup1):37–49. pmid:34872429
- 17.
Yang H, Chang F, Zhu D, Fumie M, Liu Z. Application of artificial intelligence in schizophrenia rehabilitation management: a systematic scoping review. 2024. https://arxiv.org/abs/2405.10883v2
- 18. Freeman D, Lambe S, Kabir T, Petit A, Rosebrock L, Yu LM. Automated virtual reality therapy to treat agoraphobic avoidance and distress in patients with psychosis (gameChange): a multicentre, parallel-group, single-blind, randomised, controlled trial in England with mediation and moderation analyses. Lancet Psychiatry. 2022;9(5):375–88.
- 19. Park MJ, Kim DJ, Lee U, Na EJ, Jeon HJ. A literature overview of virtual reality (VR) in treatment of psychiatric disorders: recent advances and limitations. Front Psychiatry. 2019;10:1–9.
- 20. Park K-M, Ku J, Choi S-H, Jang H-J, Park J-Y, Kim SI, et al. A virtual reality application in role-plays of social skills training for schizophrenia: a randomized, controlled trial. Psychiatry Res. 2011;189(2):166–72. pmid:21529970
- 21. Craig TK, Rus-Calafell M, Ward T, Leff JP, Huckvale M, Howarth E. AVATAR therapy for auditory verbal hallucinations in people with psychosis: a single-blind, randomised controlled trial. Lancet Psychiatry. 2018;5(1):31–40.
- 22. Rehm IC, Foenander E, Wallace K, Abbott JAM, Kyrios M, Thomas N. What role can avatars play in e-mental health interventions? Exploring new models of client-therapist interaction. Front Psychiatry. 2016;7(NOV).
- 23. Garety PA, Edwards CJ, Jafari H, Emsley R, Huckvale M, Rus-Calafell M, et al. Digital AVATAR therapy for distressing voices in psychosis: the phase 2/3 AVATAR2 trial. Nat Med. 2024;30(12):3658–68. pmid:39468363
- 24.
Stefaniak I, Sorokosz K, Janicki A, Wciórka J. Therapy based on avatar-therapist synergy for patients with chronic auditory hallucinations: A pilot study.
- 25. Østergaard SD. Will Generative Artificial Intelligence Chatbots Generate Delusions in Individuals Prone to Psychosis?. Schizophr Bull. 2023;49(6):1418–9. pmid:37625027
- 26. Khoury B, Lecomte T. Emotion Regulation and Schizophrenia. International Journal of Cognitive Therapy. 2012;5(1):67–76.
- 27. Gross JJ. Emotion Regulation: Current Status and Future Prospects. Psychological Inquiry. 2015;26(1):1–26.
- 28. Kring AM, Moran EK. Emotional response deficits in schizophrenia: insights from affective science. Schizophr Bull. 2008;34(5):819–34. pmid:18579556
- 29. Raugh IM, Strauss GP. Deconstructing emotion regulation in schizophrenia: the nature and consequences of abnormalities at the identification stage. Eur Arch Psychiatry Clin Neurosci. 2022;272(6):1061–71. pmid:34716486
- 30. Livingstone K, Harper S, Gillanders D. An exploration of emotion regulation in psychosis. Clin Psychol Psychother. 2009;16(5):418–30. pmid:19569041
- 31. Bartolomeo LA, Raugh IM, Strauss GP. Deconstructing emotion regulation in schizophrenia: The nature and consequences of abnormalities in monitoring dynamics. Schizophr Res. 2022;240:135–42.
- 32.
Beck AT. Cognitive therapy and the emotional disorders. New York, NY: International Universities Press. 1976.
- 33. Wykes T, Steel C, Everitt B, Tarrier N. Cognitive behavior therapy for schizophrenia: effect sizes, clinical models, and methodological rigor. Schizophr Bull. 2008;34(3):523–37. pmid:17962231
- 34.
Hayes SC, Strosahl KD, Wilson KG. Acceptance and commitment therapy: The process and practice of mindful change. 2nd ed. 2012.
- 35. Hayes SC. Acceptance and Commitment Therapy, Relational Frame Theory, and the Third Wave of Behavioral and Cognitive Therapies - Republished Article. Behav Ther. 2016;47(6):869–85. pmid:27993338
- 36.
Linehan MM. Cognitive-Behavioral Therapy of Borderline Personality Disorder. New York: Guilford Press. 1993.
- 37.
Segal V, Williams M, Teasdale JD. Sample chapter: Mindfulness-based cognitive therapy for depression: Second edition. 2013.
- 38. Barlow DH, Farchione TJ, Bullis JR, Gallagher MW, Murray-Latin H, Sauer-Zavala S, et al. The Unified Protocol for Transdiagnostic Treatment of Emotional Disorders Compared With Diagnosis-Specific Protocols for Anxiety Disorders: A Randomized Clinical Trial. JAMA Psychiatry. 2017;74(9):875–84. pmid:28768327
- 39. Tong A, Sainsbury P, Craig J. Consolidated criteria for reporting qualitative research (COREQ): a 32-item checklist for interviews and focus groups. Int J Qual Health Care. 2007;19(6):349–57. pmid:17872937
- 40.
World Health Organization. The ICD-10 classification of mental and behavioural disorders: clinical descriptions and diagnostic guidelines. Geneva: World Health Organization. 1992.
- 41. Kimhy D, Vakhrusheva J, Jobson-Ahmed L, Tarrier N, Malaspina D, Gross JJ. Emotion awareness and regulation in individuals with schizophrenia: Implications for social functioning. Psychiatry Res. 2012;200(2–3):193–201. pmid:22749227
- 42. Favrod J, Nguyen A, Tronche AM, Blanc O, Dubreucq J, Chereau-Boudet I, et al. Impact of Positive Emotion Regulation Training on Negative Symptoms and Social Functioning in Schizophrenia: A Field Test. Front Psychiatry. 2019.
- 43. Sadka O, Antle A. Interactive technologies for emotion regulation training: A scoping review. Int J Hum Comput Stud. 2022;168:102906.
- 44.
Bocklisch T, Faulkner J, Pawlowski N, Nichol A. Rasa: Open Source Language Understanding and Dialogue Management. 2017. https://arxiv.org/abs/1712.05181v2
- 45.
Gabor-Siatkowska K, Sowanski M, Pudo M, Rzatkiewicz R, Stefaniak I, Kozlowski M. Therapeutic Spoken Dialogue System in Clinical Settings: Initial Experiments. In: 2023.
- 46. Kozłowski M, Gabor-Siatkowska K, Stefaniak I, Sowański M, Janicki A. Enhanced Emotion and Sentiment Recognition for Empathetic Dialogue System Using Big Data and Deep Learning Methods. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). 2023. 465–80. https://link.springer.com/chapter/10.1007/978-3-031-35995-8_33
- 47. Gabor-Siatkowska K, Sowański M, Rzatkiewicz R, Stefaniak I, Kozłowski M, Janicki A. AI to Train AI: Using ChatGPT to Improve the Accuracy of a Therapeutic Dialogue System. Electronics. 2023;12(22):4694.
- 48.
Bunk T, Varshneya D, Vlasov V, Nichol A. DIET: Lightweight Language Understanding for Dialogue Systems. In: 2020. https://arxiv.org/abs/2004.09936v3
- 49. Zygadło A, Kozłowski M, Janicki A. Text-Based Emotion Recognition in English and Polish for Therapeutic Chatbot. Applied Sciences. 2021;11(21):10146.
- 50. Depow GJ, Francis Z, Inzlicht M. The Experience of Empathy in Everyday Life. Psychol Sci. 2021;32(8):1198–213. pmid:34241543
- 51. Fischer J, Dyball R, Fazey I, Gross C, Dovers S, Ehrlich PR, et al. Human behavior and sustainability. Frontiers in Ecol & Environ. 2012;10(3):153–60.
- 52. Kabat-Zinn J. Mindfulness. Mindfulness (N Y). 2015;6(6):1481–3.
- 53. Davis LL. Instrument review: Getting the most from a panel of experts. Applied Nursing Research. 1992;5(4):194–7.
- 54. Overall JE, Gorham DR. The Brief Psychiatric Rating Scale. Psychol Rep. 1962;10(3):799–812.
- 55.
Braun V, Clarke V. Thematic analysis: a practical guide. 2022.
- 56. Provoost S, Lau HM, Ruwaard J, Riper H. Embodied Conversational Agents in Clinical Psychology: A Scoping Review. J Med Internet Res. 2017;19(5):e151. pmid:28487267
- 57. Bickmore TW, Picard RW. Establishing and maintaining long-term human-computer relationships. ACM Trans Comput-Hum Interact. 2005;12(2):293–327.
- 58. Moore RK. A Bayesian explanation of the “Uncanny Valley” effect and related psychological phenomena. Sci Rep. 2012;2:864. pmid:23162690
- 59. Wang S, Lilienfeld SO, Rochat P. The Uncanny Valley: Existence and Explanations. Review of General Psychology. 2015;19(4):393–407.
- 60. Tinwell A, Grimshaw M, Nabi DA, Williams A. Facial expression of emotion and perception of the uncanny valley in virtual characters. Comput Human Behav. 2011;27(2):741–9.
- 61. Bordin ES. The generalizability of the psychoanalytic concept of the working alliance. Psychotherapy: Theory, Research & Practice. 1979;16(3):252–60.
- 62. Stein JP, Ohler P. Venturing into the uncanny valley of mind—the influence of mind attribution on the acceptance of human-like characters in a virtual reality setting. Cognition. 2017;160:43–50.
- 63. Thwaites R, Bennett-Levy J. Conceptualizing empathy in cognitive behaviour therapy: Making the implicit explicit. Behavioural and Cognitive Psychotherapy. 2007;35(5):591–612.
- 64. Safran JD, Muran JC, Eubanks-Carter C. Repairing alliance ruptures. Psychotherapy (Chic). 2011;48(1):80–7.
- 65.
Norcross JC, Lambert MJ. Psychotherapy Relationships that Work: Volume 1: Evidence-Based Therapist Contributions. Oxford University Press. 2019.
- 66. Berry N, Bucci S, Lobban F. Use of the Internet and Mobile Phones for Self-Management of Severe Mental Health Problems: Qualitative Study of Staff Views. JMIR Ment Health. 2017;4(4):e52. pmid:29092809
- 67. Garety P, Ward T, Emsley R. The role of cognitive and social factors in psychosis: a systematic review. Psychol Med. 2020;50:11–21.
- 68. Feine J, Gnewuch U, Morana S, Maedche A. A taxonomy of social cues for conversational agents. Int J Hum Comput Stud. 2019;132:138–61.
- 69. Bogie BJM, Noël C, Gu F, Nadeau S, Shvetz C, Khan H. Using virtual reality to improve verbal episodic memory in schizophrenia: A proof-of-concept trial. Schizophr Res Cogn. 2024;36:100305.
- 70. Jeppesen UN, Vernal DL, Due AS, Mariegaard LS, Pinkham AE, Austin SF, et al. Virtual reality-based versus standard cognitive behavioral therapy for paranoia in schizophrenia spectrum disorders: a randomized controlled trial. Nat Med. 2025;31(10):3425–39. pmid:40804323
- 71. Pérez-Ferrara DA, Flores-Medina GY, Landa-Ramírez E, González-Sánchez DJ, Luna-Padilla JA, Sosa-Millán AL, et al. Social cognition training using virtual reality for people with schizophrenia: a scoping review. Virtual Reality. 2024;28(2).