Through the success of chatGPT, chatbots have become increasingly available as a tool in psychotherapy. While most research focusses on the feasibility of using chatbots as therapists, we take a different approach by using chatGPT to simulate patients for psychotherapy training. We ground our chatbot design in two frameworks of therapist-patient interactions: The rupture-repair model and Kiesler’s interpersonal circumplex model. The rupture-repair model characterizes therapist-patient interactions as a cycle of breakdowns of the relationship (rupture) and the subsequent restauration of trust and collaboration (repair). Kiesler’s interpersonal circumplex model characterizes these interactions along two orthogonal dimensions: dominance-submissiveness and acceptance-rejection. Since the therapeutic alliance is one of the key factors in therapy success, it is important to incorporate hands-on practice into psychotherapy training. We propose that this can be practiced using chatbots that are prompted to correspond to different interpersonal profiles of the Kiesler model in order to create challenging therapy scenarios. Here, we present a proof-of-concept study in which we train human simulated patients and chatGPT in the same patient script. We then ask psychotherapists in training to interact with human simulated patients in a teletherapy setting. Unbeknownst to them, in half the cases the human simulated patients simply read out chatGPT responses. We use questionnaires and interviews to identify rupture-repair processes, to evaluate whether the intended interpersonal position according to the Kiesler model was achieved, and to compare the quality of the chatGPT simulated patient with the human simulated patient.