Towards scalable psychological first aid training: an autoethnographic exploration of the potential of large language models in simulation-based learning

<b>Background:</b> Psychological First Aid (PFA) is a widely used approach to provide psychosocial support to individuals in the aftermath of a stressful event. PFA training requires a costly, structured approach, including repetitive simulation-based training with actors. The rapid emergence of large language models (LLMs) could provide easy access to PFA simulation training, enabling trainees to practice conversations repeatedly and flexibly, learn from mistakes, and build emotional resilience. In addition, this would allow for training in geographically isolated regions where PFA training may be most needed.<b>Objective:</b> To explore the potential of LLMs for PFA simulation training.<b>Methods:</b> Using a digital autoethnography approach, we conducted three role plays with ChatGPT-3.5-turbo, an LLM-based chatbot, in scenarios addressing challenges typical of PFA training. In each role-play, the chatbot simulated an individual in the aftermath of a stressful situation, and a researcher took on the roles of both a trainer configuring a simulation session and a learner serving as a PFA provider. The chatbot's ability to respond, adapt, and facilitate the training process was evaluated using reflexive thematic analysis.<b>Results:</b> Our study showed that ChatGPT effectively portrayed realistic characters, collaborated effectively with the trainer, and provided feedback to the learner. However, we also noted challenges concerning overly stereotyped characterizations, identity confusion, and following the trainer's instructions.<b>Conclusions:</b> While utilizing LLMs for PFA simulation training demonstrates significant promise, achieving their full educational potential requires careful refinement, thorough empirical validation, and ongoing ethical oversight.