This exploratory study investigates the use of large language models in English for Medical Purposes (EMP) and broader ESP instruction by examining the linguistic and structural reliability of text-based medical consultation simulations. Using ChatGPT (GPT-5) with a fixed prompt, 26 simulated outpatient dialogues were generated in which the model acted as a patient. The analysis shows that the dialogues are highly regular, structurally coherent, and strongly patterned, with a clear dominance of closed questions and a limited range of recurring identities and scenarios. These findings are important for medical communication, as they suggest that LLM-generated consultations can provide stable, repeatable practice for routine interactional tasks such as history-taking, symptom elicitation, and basic diagnostic discussion. From an educational perspective, the study highlights the potential of LLMs in AI in education and autonomous learning, especially as a scaffolded resource for learners who need repeated exposure to medical discourse. At the same time, the limited variability of the interactions indicates that such simulations are best used as a supplementary tool rather than a replacement for human-mediated communication practice. The study contributes to current discussion on the pedagogical value and limitations of large language models in English for Medical Purposes, ESP, and digital language learning.
Evgeni Stanchev (Thu,) studied this question.