Fostering Reflection: Development and Initial Evaluation of a Mentalization-Based Genai Simulator for Foster Care Supervisors

Background and Purpose: Foster care supervisors, primarily social workers, must navigate between the diverse and often competing needs of children, foster families, and biological parents. This role demands advanced interpersonal skills, including the capacity to maintain empathy and sensitivity toward multiple parties simultaneously—a competency rooted in reflective functioning (RF) and mentalization. Training such complex relational skills is inherently difficult: traditional approaches are resource-intensive, require skilled trainers, and offer limited opportunities for repeated, private practice with feedback. Generative AI (GenAI) offers a promising alternative by providing accessible, nonjudgmental simulated environments. This study describes the development of a GenAI-based chatbot simulator designed to enhance RF among foster care supervisors and reports on a preliminary evaluation of its usability and acceptability.

Methods: Using a mixed-methods design, 46 female foster care supervisors (Mean age = 41.6, SD = 8.8; Mean foster care experience = 8.9 years) participated in a single training session. The simulator, deployed on the PMFM platform and powered by Claude Sonnet 3.5 V2, utilizes a multi-layered prompt architecture exceeding 5,000 words to enact a realistic foster mother persona with defensive psychological patterns. The interaction included: (1) a 20-turn dialogue with “Julie,” (2) real-time scaffolding from a virtual mentor (“Dr. Menton”), and (3) a two-part feedback system revealing the bot’s “internal voice” and structured professional analysis. Usability and acceptability were assessed via seven items on a 7-point Likert scale measuring perceived utility and a recommendation item. Qualitative data from four open-ended questions were analyzed via thematic analysis.

Results: Quantitative findings indicated high usability and acceptability. One-sample tests against the scale midpoint showed that participants rated the feedback as highly contributory (M = 5.83, SD = 1.40; d = 1.31, p < .001), with 67.4% assigning scores of 6 or 7. Participants also reported gains in practical skills for supporting foster parents (M = 5.50, SD = 1.21) and increased confidence in managing work-related problems (M = 5.24, SD = 1.29). Willingness to recommend the tool was exceptionally strong (M = 6.67, SD = 0.77; d = 3.47, p < .001), with 93.4% endorsing high ratings. Qualitative thematic analysis yielded four themes: (1) a dialectic of realism—valuing authentic dialogue while desiring greater clinical resistance; (2) the simulator as a reflective space enabling deliberate, unhurried professional practice; (3) the pedagogical scaffold’s dual role as guiding compass and potential source of overload; and (4) complex reactions to the AI “other,” ranging from awe to apprehension about implications for relational work.

Conclusions and Implications: This study demonstrates how GenAI-based simulation can serve as a scalable, low-cost tool for training reflective functioning in social work practice. By providing a safe environment for repeated practice without fear of judgment, the simulator offers an accessible complement to traditional supervision that can potentially be adapted across languages and cultural contexts. Ethical implementation requires ongoing attention to algorithmic bias and ensuring that AI remains a tool to amplify, rather than replace, human relational expertise.