Background: Effective physician–patient communication is a core competency in undergraduate medical education, yet students have limited real patient contact and mostly practice with peers or standardized patients (SPs). SPs are effective but resource intensive and may provoke performance anxiety. Generative AI (GenAI)–based virtual patients may offer a scalable, low threshold way to rehearse basic communication skills before high stakes encounters with SPs.
Intervention: The University of Amsterdam developed “pAItient sim”, a GenAI virtual patient that enables students to practice communication in a text based chat. The system uses GPT 4.1 on a secure institutional platform, constrained by a behavioural prompt and structured case descriptions to ensure medically plausible, role consistent and didactically useful responses. Earlier small scale pilots suggested enthusiasm and perceived usefulness, but learning impact remained unclear.
Methods: We conducted an observational mixed methods study with second year medical students, for whom pAItient sim was offered as optional preparation alongside existing history taking training. We examined: (1) how students communicate with the virtual patient, (2) how realistic and didactically appropriate its behaviour is, and (3) how, how often and why students choose to use or not use the tool. Data sources were anonymized chat transcripts, usage analytics and a post class questionnaire; transcripts were analysed using directed qualitative content analysis.
Results: Of 248 students, 52 used pAItient sim, generating 99 transcripts. Most student questions were biomedical, with relatively few addressing patient perspectives or expressing empathy, matching what students reported they saw as the tool’s main purpose. Non users mainly cited limited time and doubts about the realism and added value of a text based GenAI chatbot.
Conclusion: We offer design implications for GenAI based virtual patients, discuss usage patterns and identify factors that support or hinder student use.