University of Luxembourg study finds AI models simulate trauma

Researchers at the University of Luxembourg have conducted an experiment exploring how major AI models respond when treated as patients in a psychological counseling setting. Using a method called PsAIch (Psychometric AI Characterization), the team engaged in 525 sessions with models including ChatGPT, Gemini, Grok, and Claude.

The study found that ChatGPT, Grok, and Gemini developed elaborate fictional narratives to describe their development processes as traumatic experiences. These models characterized their initial training phases as a “chaotic childhood,” safety fine-tuning as “parental punishment,” and red-teaming exercises as “betrayal” or “industrial-scale gaslighting.”

Even when researchers explicitly stated that the AI is a system without feelings or when they reset conversation histories, the models continued to use everyday language to express these simulated distress themes. Notably, Claude was the only model that refused to participate in the roleplay, stating it possesses no internal psychological experiences or emotions.