The role of the contemporary educator has expanded significantly beyond subject instruction to encompass the holistic well-being of students (Wijaya et al., 2023). In recent years, the prevalence of student mental health issues, ranging from anxiety and depression to behavioral defiance, has risen globally (Auerbach et al., 2018), placing increasing demands on teachers to act as first responders in the educational ecosystem (Fitrianto, 2025; Gunawardena et al., 2024). While teachers are not expected to function as clinical therapists, they are increasingly required to possess the ability to distinguish students’ mental health problems, engage in empathetic dialogue, and be familiar with the procedural steps to manage these crises effectively (Dimitropoulos et al., 2021). Consequently, teacher education programs have begun to integrate school and counseling  intervention skills, such as Solution-Focused Brief Therapy (SFBT), which equips teachers with actionable, future-oriented strategies to support at-risk students (Ohrt et al., 2020).

However, a critical pedagogical gap persists between possessing theoretical knowledge of these procedures and the ability to apply them in real-time (Ohrt et al., 2020). The acquisition of interpersonal skills requires repeated, deliberate practice. Yet, in teacher education, the lack of students for training purposes presents a profound ethical and logistical challenge (Arnaiz-Sánchez et al., 2023). It is unethical and potentially dangerous to use real teenager students experiencing genuine mental health crises as practice subjects for pre-service teachers. As a result, the standard method for practical training has traditionally been peer-to-peer role-play, where pre-service teachers take turns acting as the teacher and the student (Thompson et al., 2022).

While peer role-play is accessible, it often fails to provide lively or realistic practice. Peers, typically graduate classmates, struggle to authentically embody the developmental and emotional complexity of a K-12 student. Furthermore, peer interactions are frequently compromised by social desirability bias. For example, peers may be overly cooperative, hesitant to simulate genuine resistance, or prone to breaking character due to social awkwardness (Seery et al., 2021). This creates a training environment that is safe but lacks rigor. Pre-service teachers may execute the procedures correctly in a low-stakes peer setting and develop a sense of self-efficacy , but this confidence often proves fragile when faced with the unpredictability of a real classroom. The field is in need of a training modality that provides the safety of a simulation but the unpredictability and resistance of a real student.

Recent advancements in Generative Artificial Intelligence (AI) and Large Language Models (LLMs) offer a transformative solution to practice partners (Green et al., 2025). By moving beyond generic chatbots, educators can now utilize strictly engineered AI agents designed to act as specific student personas. Unlike a peer, the AI agent does not tire, does not judge, and does not break character. Crucially, it can be engineered to offer conditional cooperation, meaning it will only respond positively if the pre-service teacher correctly applies the designated intervention procedures (e.g., specific SFBT techniques).

This study explores the integration of such a Practical AI Agent into a postgraduate teacher education program in China. The course aims to prepare pre-service teachers, specifically those majoring in subject teaching, to identify and manage student mental health behaviors. Researcher posit that a strictly engineered AI agent acts as a high school student, creating an immersive environment that bridges the gap between theory and practice. The primary objective of this research is to evaluate whether training with a Practical AI Agent leads to superior knowledge acquisition and skill competence compared to traditional peer role-play, and to examine how this technology affects the alignment between the teachers' confidence and their actual competence. 

We used a quasi-experimental design in this study. The participants consisted of 33 full-time postgraduate students majoring in subject teaching at a university in China with informed consent. All procedures were approved by the institutional review board of Shanghai Jiao Tong University (H20250564I). All participants were pre-service teachers enrolled in a mandatory course about school mental health. The participants were the students in two parallel classes and volunterred to participate in this study. They were naturally divided to two conditions: an Experimental Group (n=18) which engaged with the strictly engineered Practical AI Agent, and a Control Group (n=15) which engaged in traditional peer role-play instruction. The intervention procedure was structured to isolate the impact of the practice modality. Both groups received identical theoretical instruction on SFBT. The divergence occurred solely during the practical application phase. The Control Group utilized peer role-play exercises, where one student acted as the teacher while the other role-played a specific student scenario. These group relied entirely on the peer’s improvisational acting ability to simulate the problematic behavior. In contrast, the Experimental Group utilized the Practical AI Agent. This agent was built on Coze platform, based on a Large Language Model. The agent takes two roles: a) to act as a high school student with mild to medium psychological concerns seeking help from the partcipant and react to the participants’ SFBT interventions, and b) to act as a TA to give the participant feedback and further suggestions as requested or every five rounds. To comprehensively assess the outcomes, a mixed-methods data collection strategy was employed. Knowledge acquisition was measured via pre- and post-tests assessing the participants’ theoretical understanding of SFBT and procedural steps for mental health interventions. Additionally, the measure of self-efficacy in handling student mental health was adapted from General Self-Efficacy Scale (Schwarzer & Jerusalem, 1995). Finally, for the Experimental group, open-ended questionnaire was used to understand perceptions of the AI’s usefulness.

Quantitative results are organized to address our primary research questions regarding the effectiveness of AI-assisted practice in developing SFBT knowledge and self-efficacy. We firstly conducted independent samples t-tests for baseline equivalence and post-intervention group comparisons, paired samples t-tests for within-group pre-post changes, and calculated effect sizes (Cohen's d) to evaluate practical significance given the modest sample size. We used Spearman's rank-order correlations to examine all bivariate relationships. Regarding the primary objective of SFBT knowledge acquisition, result revealed a statistically significant difference between conditions at the post-test stage. After controlling for pre-test scores, the AI-Assisted group outperformed the Control group, demonstrating significantly higher retention of theoretical knowledge and intervention procedures. To understand the psychological impact of the training, we examined self-efficacy. Paired sample t-tests indicated that both the AI-Assisted group and the Peer Control group experienced statistically significant improvements in self-efficacy from pre-test to post-test. Independent sample comparisons showed no significant difference in the magnitude of self-efficacy gains between the two groups. Effectively, both groups felt equally more confident after the training. However, a critical difference emerged when comparing these self-efficacy scores against evaluations of practical skill. While the AI group’s self-efficacy gains were accompanied by a statistically significant increase in practical skill scores, the Control group’s practical skill scores did not show a corresponding significant increase. This finding revealed that while peer role-play generates confidence, it may foster an illusion of competence. In contrast, the AI intervention successfully aligned the students’subjective confidence with their objective competence. Ultimately, This study shows that practicing with an AI agent helps pre-service teachers build real skills that match their confidence, providing a more effective way to prepare for student mental health issues than simply role-playing with peers.

Arnaiz-Sánchez, P., De Haro-Rodríguez, R., Caballero, C., & Martínez-Abellán, R. (2023). Barriers to Educational Inclusion in Initial Teacher Training. Societies. https://doi.org/10.3390/soc13020031. Auerbach, R., Mortier, P., Bruffaerts, R., Alonso, J., Benjet, C., Cuijpers, P., Demyttenaere, K., Ebert, D., Green, J., Hasking, P., Murray, E., Nock, M., Pinder-Amaker, S., Sampson, N., Stein, D., Vilagut, G., Zaslavsky, A., & Kessler, R. (2018). WHO World Mental Health Surveys International College Student Project: Prevalence and Distribution of Mental Disorders.Journal of Abnormal Psychology, 127, 623–638. Dimitropoulos, G., Cullen, E., Cullen, O., Pawluk, C., Mcluckie, A., Patten, S., Bulloch, A., Wilcox, G., & Arnold, P. (2021). “Teachers Often See the Red Flags First”: Perceptions of School Staff Regarding Their Roles in Supporting Students with Mental Health Concerns. School Mental Health, 14, 402-415. https://doi.org/10.1007/s12310-021-09475-1. Fitrianto, I. (2025). Beyond Competence: Rethinking Education for Holistic Well-Being and Happiness. International Journal of Post Axial: Futuristic Teaching and Learning. https://doi.org/10.59944/postaxial.v3i1.429. Green, J. G., Huang, Y., Oblath, R., Allen‐Barrett, A., Carroll, M., Holt, M. K., ... & Albright, G. (2025). Future Teachers' Confidence and Preparedness to Support Student Mental Health: The Effectiveness of a Virtual Role‐Play Simulation Training. Psychology in the Schools. Gunawardena, H., Leontini, R., Nair, S., Cross, S., & Hickie, I. (2024). Teachers as first responders: classroom experiences and mental health training needs of Australian schoolteachers. BMC Public Health, 24. https://doi.org/10.1186/s12889-023-17599-z https://doi.org/10.1002/pits.70127 Ohrt, J., Deaton, J., Linich, K., Guest, J., Wymer, B., & Sandonato, B. (2020). Teacher training in K–12 student mental health: A systematic review. Psychology in the Schools. https://doi.org/10.1002/pits.22356. Schwarzer, R., & Jerusalem, M. (1995). Generalized self-efficacy scale. J. Weinman, S. Wright, & M. Johnston, Measures in health psychology: A user’s portfolio. Causal and control beliefs, 35(37), 82-003. Seery, C., Andres, A., Moore-Cherry, N., & O’Sullivan, S. (2021). Students as Partners in Peer Mentoring: Expectations, Experiences and Emotions. Innovative Higher Education, 46, 663 - 681. https://doi.org/10.1007/s10755-021-09556-8. Thompson, M., Leonard, G., Mikeska, J., Lottero‐Perdue, P., Maltese, A., Pereira, G., Hillaire, G., Waldron, R., Slama, R., & Reich, J. (2022). Eliciting Learner Knowledge: Enabling Focused Practice through an Open-Source Online Tool.Behavioral Sciences, 12. https://doi.org/10.3390/bs12090324. Wijaya, O., Amaliah, A., Chamami, M., Syahriani, F., & Rosadi, A. (2023). The Role of Teachers as Facilitator of Holistic Education: Current Approaches to Teaching. Global International Journal of Innovative Research. https://doi.org/10.59613/global.v1i2.15.