From Simulation to Real Therapy: How Do We Know Training Works?

A trainee completes ten simulated sessions. They feel more confident, receive better automated scores and describe the experience as useful. But have they become more skilled as a therapist?

This question matters for every training institution considering AI. Enjoyable practice and demonstrable learning are not interchangeable.

Three different outcomes

User experience: Learners find the tool engaging, easy to use and comfortable for practising mistakes. That matters for adoption, but it does not establish skill acquisition.

Performance in simulation: Across practice scenarios, a learner may become better at naming emotions, asking open questions or avoiding premature advice. Even then, we must ask how performance was assessed and who set the criteria.

Transfer: The learner responds more effectively in a new situation, perhaps with a human standardized client and an independent assessor. This outcome comes closer to the central educational aim.

What does recent research show?

A 2026 randomized study involving 87 participants compared AI-enabled virtual patient sessions with prerecorded role-play videos. Virtual-patient training improved perceived competence and self-efficacy and reduced insecurity, but it did not outperform video training overall or improve knowledge. These findings are promising for a complementary training tool. They do not demonstrate transfer to actual clinical encounters.

How could we test transfer more carefully?

Consider a programme focused on emotional reflection. A practical evaluation might involve:

1. Baseline task: A short common scenario before training.
2. Focused practice: Varied situations with explicit feedback.
3. Fresh post-training task: A new scenario rather than an identical replay, reducing the influence of memorization.
4. Independent assessment: A shared rubric, ideally with assessors blinded to whether the task was performed before or after training.

A follow-up several weeks later can examine retention. Demonstrating improvement with real clients requires additional study, appropriate supervision and careful ethical safeguards.

Two traps to avoid

Higher scores from the same AI system that generated the patient do not independently certify competence. A learner may simply have adapted to that system's preferences. Likewise, confidence is worth measuring, but it is not the same as observed clinical skill.

The productive question is not merely, “How many simulations did learners complete?” It is, “Which skill improved, how was it measured, and did that improvement appear in a new setting?”

From theory to practice: TheraMentor offers opportunities for focused simulation and reflection. Decisions about actual clinical competence belong with qualified human educators and supervisors.