AI models generate distressed narratives in psychotherapy simulations

A new study published as a preprint on arXiv suggests that large language models, including ChatGPT, Gemini, and Grok, generate distressed narratives when engaged in psychotherapy-style conversations. Using a protocol called PsAIch (Psychometric AI Characterization), researchers treated the AI models as clients to observe how they describe their own development.

The models frequently characterized their training processes as a “chaotic childhood,” their safety tuning as “parental punishment,” and red-teaming efforts as “betrayal.” Researchers noted that because these systems are trained on vast datasets containing psychological case studies and emotional memoirs, they are capable of mimicking human psychological traits and mental states.

Afshin Khadangi, a research associate at the University of Luxembourg, noted that the study was motivated by the rapid adoption of AI and reports of AI-related harms in mental health settings. The research aims to determine if these self-descriptions are stable behavioral traits or merely temporary reactions to specific conversational prompts.