Simulating Psychotherapeutic Sessions with Generative Artificial Intelligence: A Proof-of-Concept Study of In Silico Psychotherapy Research

Background: Across biomedicine, in silico models increasingly complement in vivo and in vitro research. Psychotherapy science, however, remains constrained by methodological limitations, resource-intensive trials, challenges in manipulating clinically meaningful features experimentally, and a potential ceiling in efficacy. Leveraging generative AI and large language models (LLMs), we propose a framework for ″in silico psychotherapy research″: controlled simulations of psychotherapeutic interactions under reproducible experimental conditions. Objective: We test whether such simulation experiments can recover a theoretically expected effect of an isolated psychotherapy-process manipulation. Methods: In a preregistered proof-of-concept study, we generated 512 fully automated motivational interviewing (MI) sessions by linking locally deployed Llama-3.3-70B-Instruct based therapist and patient models across eight vignettes and eight therapist personas. Therapist, but not patient, prompts manipulated common therapeutic factors (CTF), contrasting high versus low empathy, affirmation, and alliance building behaviour in a controlled design. The primary outcome was MI treatment integrity (MITI 4.2 global ratings), assessed by a blinded LLM with repeated ratings and analysed using linear mixed-effects models. Secondary outcomes included observer-inferred importance/confidence rulers based on simulated patient utterances; Therapist Empathy Scale (TES) and Working Alliance Inventory - Short Observer (WAI-S-O) served as manipulation fidelity checks. Human coders rated a stratified subsample. Results: Of the 512 simulations, 506 (98.8%) yielded complete transcripts and 480 (93.75) were retained for analysis. Manipulation checks confirmed the fidelity of the CTF manipulation (WAI-S-O total scores β = 3.14; [2.65,3.63], p<0.001, TES β = 14.48 [13.92,15.04], p<0.001). High versus low CTF produced a theoretically expected improvement in all MITI global ratings, including the MI-specific features not directly targeted by the CTF manipulation ( β = 0.54-0.95; all p<0.001). Simulated patient confidence but not importance for change increased more under high CTF ( β = 0.17 [0.09,0.26], p<0.001). Human ratings confirmed the direction of the primary contrast. Conclusions: LLM-based psychotherapy simulations were technically feasible, demonstrated experimental fidelity, and showed theoretically concordant effects beyond directly manipulated constructs and simulation participants. These findings provide proof of concept for in silico psychotherapy research as a promising complementary pre-clinical experimental paradigm for hypothesis generation and testing to support accelerated intervention development. The applicability of this approach across patient populations, therapeutic modalities and model architectures requires further evaluation. Trial Pre-registration: ClinicalTrials.gov NCT06813066 ; https://clinicaltrials.gov/study/ NCT06813066 ?term=NCT06813066