Therapy sessions can now run in silico: raising therapist empathy prompts lifts every motivational-interviewing integrity score

Psychotherapy sessions can be generated at scale with no human participants and used as controlled experiments, with large language models playing both therapist and patient across a whole session. Psychotherapy research has relied on resource-intensive human trials in which clinically meaningful features are hard to manipulate precisely or randomise. Of 512 simulated sessions, 506 (98.8%) produced complete transcripts; 26 were excluded for severe simulation errors such as looping or role deviation, leaving 480 in the analysis set.

Changing only the degree of empathy, affirmation and alliance building in the therapist prompt makes sessions score systematically better or worse on motivational interviewing ratings. The manipulation targets relational stance, yet the ratings include motivational-interviewing-specific dimensions such as cultivating change talk and softening sustain talk that were not directly manipulated. Linear mixed-effects models with random intercepts for patient vignette and therapist persona gave adjusted mean differences between high and low common factors of 0.54 to 0.95 points on the four motivational interviewing treatment integrity global ratings on a 1 to 5 scale, all p<0.001.

Simulated patients also respond selectively to therapist stance, with a larger observer-inferred gain in confidence for change under high common factors. The patient-side prompts were identical across conditions and were not manipulated. The pre-to-post change in the confidence ruler was higher under high common factors, beta=0.17 (95% CI 0.09 to 0.26, p<0.001), while the importance ruler showed no clear condition difference.

Manipulation checks confirm the therapists actually behaved as instructed, and human raters reproduced the direction of the primary contrast. Generated text alone cannot show whether the model followed the prompt, so independent fidelity checks are needed. Working Alliance Inventory Short Observer total beta=3.14 (95% CI 2.65 to 3.63) and Therapist Empathy Scale beta=14.48 (95% CI 13.92 to 15.04), both p<0.001; a stratified subsample of 40 simulated protocols plus 8 real transcripts was rated blind by two raters each, and human ratings likewise placed high common factors above low on motivational interviewing global ratings and key relational measures.

The demonstrated scope is 480 analysable simulations on a locally deployed Llama-3.3-70B-Instruct, using motivational interviewing as the single therapy, eight patient vignettes and eight therapist personas, in single written sessions, reproducing the expected direction of the therapist-side manipulation. The authors position the paradigm as a preclinical research stage for testing process theories and measurement strategies, not as a substitute for clinical evidence from patient outcomes.

The authors report that simulated sessions were rated markedly less credible than real transcripts for both therapist and patient behaviour. Simulations scored below real sessions on the empathy scale and the alliance bond subscale but higher on the task and goal subscales. Agreement between model and human ratings varied by outcome, stronger for the empathy total and weaker for credibility and some motivational interviewing talk-related constructs. Shared bias between simulating and evaluating models could also create evaluation blind spots. Simulations were limited to single written sessions without non-verbal cues or session-to-session dynamics, and patient change was an observer-inferred indicator from utterances rather than a clinical effect.