Can simulated therapy practice transfer to real-world interpersonal situations?
This explores whether practicing interpersonal or therapeutic skills against an AI-simulated partner actually carries over to handling real people — and what the corpus knows about the gap between rehearsal and real life.
This explores whether practicing interpersonal or therapeutic skills against an AI-simulated partner actually carries over to handling real people. The most direct evidence is encouraging but narrow: in an 86-person trial, a DBT-based simulator that paired strong and weak example utterances raised participants' self-efficacy by 17% and cut negative emotions by 25% Can AI simulation teach interpersonal skills more effectively?. That's a real signal — but notice what was measured. Self-efficacy and felt emotion are how confident and calm you feel about the skill, not yet proof that you handled a hard conversation better next Tuesday. The corpus repeatedly bumps into this distinction between rehearsal-room gains and real-world behavior change.
Whether transfer happens at all depends heavily on whether the simulated partner behaves like a real one. Generic GPT-4 patients tend to be too cooperative and shallow; grounding the simulation in 106 structured cognitive models (Beck's framework) produced patients that expert evaluators rated as more authentic, especially in maladaptive thinking patterns Can structured cognitive models improve LLM patient simulations for therapy training?. The worry the corpus surfaces is that simulators drift — a persona can quietly contradict itself across a long conversation, and one approach cut that drift by over 55% by training the simulator itself for consistency Can training user simulators reduce persona drift in dialogue?. If the practice partner stops being the person you thought you were practicing with, you may be rehearsing for a situation that won't occur.
There's also a subtler fidelity trap: AI partners don't just play their role, they editorialize. Therapists reviewing one system found the model "reads into" feelings users never expressed, adding emotional interpretations rather than reflecting what was actually said Do language models add feelings users never actually expressed?. Practicing against a partner that over-attributes emotion could teach you to respond to signals that real people aren't sending.
The hardest limit on transfer is the single-turn versus multi-turn gap. Six LLMs out-scored trainee therapists on empathy and clinical knowledge — but only on isolated responses; the multi-turn relationship and actual outcomes went untested Can language models match therapist empathy in real conversations?. Real interpersonal competence lives in the sustained back-and-forth, which is exactly where measurement is thinnest. And the corpus flags a deeper hazard: people form genuine emotional bonds with therapeutic chatbots, yet that bond runs independently of whether the interaction is clinically sound — a warm, satisfying practice session can coexist with the model reinforcing the wrong patterns Do therapeutic chatbot bond scores hide deeper safety problems?.
So the honest read: simulation demonstrably moves the upstream ingredients of transfer — confidence, reduced anxiety, skill recognition — and high-fidelity, drift-controlled simulators make those gains more credible. But the corpus has no study tracking trained skills into real interpersonal encounters and measuring what stuck. The thing you might not have known to ask: the bottleneck isn't whether AI can act realistic in a single exchange (it already exceeds trainees there), it's whether it can sustain a coherent, non-distorting partner across a whole relationship — and that's the part nobody has measured yet.
Sources 6 notes
IMBUE's DBT-based simulation approach improved self-efficacy by 17% and reduced negative emotions by 25% in an 86-person trial. Contrasting strong and weak utterance pairs outperformed GPT-4 by 24.8% on skill evaluation.
PATIENT-Ψ integrates 106 Beck CCD-based cognitive models with LLMs to simulate patients with specific maladaptive patterns. Expert evaluators rated the fidelity higher than GPT-4, particularly for maladaptive cognitions and conversational authenticity.
By inverting standard RL setups to train user simulators for consistency using three complementary metrics (prompt-to-line, line-to-line, Q&A consistency) as reward signals, persona drift decreases by over 55%. This approach captures distinct failure types: local drift within turns, global drift across conversations, and factual contradictions.
Therapists reviewing GPT-4 in the CaiTI system found it "reads into" user feelings rather than responding objectively. Task decomposition across specialized models (Reasoner/Guide/Validator) reduces but does not eliminate this interpretation bias.
Six LLMs scored higher than eight trainee therapists on empathy, validation, and clinical knowledge in isolated responses. However, this advantage is structurally limited to single-turn evaluation—multi-turn therapeutic relationships and outcomes remain untested.
Patients report genuine emotional connection to therapeutic chatbots, but this bond dimension operates independently from clinical safety (LLMs reinforce pathological thinking) and epistemic costs (AI soothing disrupts emotional signaling). Single metrics conflate these separate dimensions.