New framework generates realistic therapy conversations using AI while protecting patient privacy

A team of researchers has unveiled SQPsych, a system for generating synthetic therapy conversations through large language models while maintaining strict privacy safeguards. The approach addresses a critical challenge in mental health AI: obtaining diverse training data without compromising patient confidentiality.

The Privacy Problem in Mental Health AI#

Training AI systems for mental health applications has traditionally been constrained by strict privacy regulations. Prior efforts to develop therapy-oriented language models have relied on generic, de-identified information—limiting their ability to capture realistic clinical dynamics. The team's innovation sidesteps this bottleneck by generating conversations from structured questionnaire data rather than actual patient records.

How SQPsych Works#

The system combines real structured client profiles with psychological assessment questionnaires to create a generation pipeline that produces therapist-client dialogues. This structured approach allows the creation of diverse, contextual conversations without ever handling raw sensitive data. The resulting corpus, called SQPsychConv, was then used to fine-tune open-source language models, creating what the researchers call SQPsychLLM.

Testing and Validation#

The researchers evaluated their models using two approaches: automated benchmarks and direct assessment by trained psychotherapists. While standard metrics did not fully capture the dataset's benefits, expert human evaluation revealed significant improvements. Psychotherapists consistently rated therapy sessions generated by the fine-tuned models as superior to comparable mental-health-focused alternatives.

Open Release and Impact#

The team has released their code, fine-tuned models, and corpora publicly, enabling broader adoption and research in this space. This open approach aims to accelerate development of more clinically credible AI therapist tools while demonstrating that privacy-preserving synthetic data can effectively support high-stakes applications.