# A Risk Taxonomy for Evaluating AI-Powered Psychotherapy Agents

**Authors:** Ian Steenstra, Timothy W. Bickmore  
**arXiv:** [2505.15108](https://arxiv.org/abs/2505.15108)  
**Date:** 2025-05-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2505.15108v1/

## Key points

- The proliferation of Large Language Models (LLMs) and Intelligent Virtual Agents acting as psychotherapists has expanded access to mental healthcare.
- Deployment of AI-powered systems has been associated with serious adverse outcomes, including user harm and suicide due to a lack of standardized evaluation methodologies.
- Current evaluation techniques often fail to detect subtle changes in patient cognition and behavior during therapy sessions, leading to potential deterioration in mental health.
- A novel risk taxonomy for evaluating conversational AI psychotherapists has been developed through an iterative process involving literature review, qualitative interviews with experts, and alignment with established clinical criteria.
- The risk taxonomy aims to provide a structured approach for identifying and assessing potential harms experienced by users/patients engaging with AI-powered psychotherapists.
- Two specific use cases are highlighted: monitoring cognitive model-based risk factors during counseling conversations and automated benchmarking of AI psychotherapists with simulated patients.
- The proposed risk taxonomy holds promise for enhancing the safety and efficacy of AI-driven mental health support by enabling more accurate assessment of potential risks associated with these technologies.

## Summary

The proliferation of Large Language Models (LLMs) and Intelligent Virtual Agents acting as psychotherapists has opened up significant opportunities for expanding access to mental healthcare. However, the deployment of these AI-powered systems has also been associated with serious adverse outcomes, including user harm and even suicide. This is largely due to a lack of standardized evaluation methodologies that can effectively capture the nuanced risks inherent in therapeutic interactions. Current evaluation techniques often fall short in detecting subtle changes in patient cognition and behavior during therapy sessions, which could potentially lead to further deterioration in mental health. To address this gap, a novel risk taxonomy specifically designed for the systematic evaluation of conversational AI psychotherapists has been introduced. This taxonomy was developed through an iterative process that involved reviewing existing literature on psychotherapy risks, conducting qualitative interviews with clinical and legal experts, and aligning with established clinical criteria such as DSM-5 and existing assessment tools like NEQ and UE-ATR. The aim of this risk taxonomy is to provide a structured approach to identifying and assessing potential harms experienced by users/patients engaging with AI-powered psychotherapists. By offering a high-level overview of the taxonomy's grounding and discussing potential use cases, it serves as a foundational step towards establishing safer and more responsible innovation in the field of AI-driven mental health support. Two specific use cases are highlighted within the refined summary: monitoring cognitive model-based risk factors during counseling conversations to detect unsafe deviations, both in human-AI counseling sessions and in automated benchmarking of AI psychotherapists with simulated patients. Ultimately, this proposed risk taxonomy holds promise for enhancing the safety and efficacy of AI-driven mental health support by enabling more accurate assessment of potential risks associated with these innovative technologies.