Our work investigates the use of LLMs (Large Language Models) as training tools for prospective primary school teachers in the field of mathematics education and artificial intelligence (AI) literacy. Building on previous classroom experience of introducing AI concepts through unplugged activities, this preliminary study explores whether LLMs can simulate students, particularly those with learning difficulties, to support teacher preparation also in inclusive contexts. Two models, ChatGPT-5 and Perplexity Pro, were tested using role-play prompts designed to generate responses resembling those of real pupils. The results indicate that while the mistakes made by the artificial students do not completely overlap those observed in actual classrooms, the simulations still provide valuable insights for teacher reflection. In particular, the experiments highlighted challenges related to negation, set representation, and the didactic contract, with the artificial students sometimes reproducing behaviors documented in mathematical education research, such as anxiety-driven overgeneralization. These findings suggest that LLMs can serve as a useful resource for developing teacher awareness of potential learning difficulties and for designing strategies to address them. In the future, testing can be extended to a larger number of models and tasks, and simulations can be compared with a larger amount of empirical data from students with formally diagnosed learning disorders, particularly dyscalculia.