AI in Psychotherapy Training: What Does the Evidence Show in 2026?

Can a conversation with a virtual client help someone practise therapeutic skills? The question matters to universities, training institutes, and clinicians seeking more opportunities for structured rehearsal. Artificial intelligence (AI) is being explored for simulated sessions, feedback, and reflective learning—not simply for generating text.

The useful question is not whether AI can “be a therapist.” It is which educational tasks AI can support, how those tasks should be evaluated, and where the evidence stops.

Three ways AI enters therapist training

A simulated client allows a learner to practise questions, reflections, and responses in a fictional encounter. Automated feedback may highlight observable communication patterns. Repeatable exercises can help learners return to the same skill across different scenarios, rather than hoping it appears during occasional role-play sessions.

These functions should not be confused. Conversational realism is not the same as clinically valid assessment. Similarly, greater learner confidence does not necessarily mean better outcomes with real clients.

What has research found?

In 2024, researchers introduced PATIENT-ψ, an LLM-based client simulation framework for cognitive behavioural therapy (CBT) training. A user study with 13 mental health trainees and 20 experts reported promising perceptions of skills learning and realism. This was an early evaluation, not evidence of long-term clinical effectiveness.

A 2026 randomized study involved 87 psychology students and psychotherapy trainees. Participants completed four AI virtual-patient sessions with feedback or watched recorded role-play videos. Those using simulations reported improvements in perceived competence and self-efficacy, alongside reduced insecurity. Importantly, the simulated-client condition did not outperform video training overall, and knowledge did not improve. That nuance matters when describing what AI training can deliver.

Qualitative research has also identified a recurring challenge: AI clients may be too agreeable, emotionally flat, or culturally generic unless carefully designed. These limitations can make practising with them different from meeting a real person.

What makes training meaningful?

A helpful exercise starts with a clear target: perhaps asking fewer leading questions or accurately reflecting emotion. The learner tries a short scenario, reviews a specific response, and reflects on what might be different next time. In an educational setting, a supervisor can place that experience in a wider clinical framework.

Without a goal or reflection, a fluent conversation may simply feel productive without demonstrating any meaningful learning.

Can AI replace clinical supervision?

Current evidence does not support that conclusion. AI-supported practice may be a useful addition to peer role-play, skills workshops, supervision, and supervised clinical experience. It cannot fully reproduce the relational, ethical, and contextual complexity of real practice, and AI feedback can be wrong.

TheraMentor is designed around that narrower educational purpose: trying responses in simulated sessions and reflecting on decisions before they matter in a real therapeutic setting. It is not therapy and does not replace professional supervision.

From theory to practice: explore simulated skills training with TheraMentor.