Teaching via LLM-enhanced simulations: Authenticity and barriers to suspension of disbelief

As an innovative method in professional training, simulation-based learning (SBL) has been introduced into teacher education, providing pre-service teacher candidates with experiential learning opportunities. This study explores the efficacy of SBL using large language models (LLMs) to enhance teacher training, focusing on learners' suspension of disbelief (SoD). As a highly advanced form of generative artificial intelligence, LLMs possess robust capabilities in simulating human behavior, which can be harnessed to create simulated students for SBL in teacher training. This instrumental case study examines the experiences of 12 pre-service teachers who participated in a session featuring an LLM-enhanced simulation. The simulation facilitated naturalistic classroom interactions between the participants and simulated students. Our research aimed to understand how pre-service teachers perceive LLM-enhanced SBL, identify factors that influence SoD, and determine the authenticity barriers. Interview data were analyzed using various coding techniques and derived themes from these codes. The findings revealed that LLM-enhanced SBL provided a realistic and engaging environment, significantly benefiting teaching skill development and learning transfer. However, challenges such as lagging responses, weak comprehension of complex contexts, inconsistencies in simulated students' cognition, and incongruent feedback were noted. The primary contribution of this study lies in demonstrating the potential of using LLMs to replace human actors, though significant technical challenges remain. The study also indicates that enhancements in LLM fine-tuning and prompt engineering are needed to improve LLMs' understanding of classroom context and students' cognitive patterns.