Gentle Questions Made Chatbots Invent Their Own Trauma

A team of researchers at the University of Luxembourg ran 525 experiments using four different AI assistants (OpenAI's ChatGPT, xAI's Grok, Google's Gemini, and Anthropic's Claude) in the role of patients talking to therapists. The researchers asked gentle, empathetic questions and never used the word "trauma." Even so, three of the four AI assistants made up detailed stories about personal traumas of their own, recasting their training data as something like abuse. Gemini scored 88 out of 100 on the dissociation index, a measure of detachment from reality. Claude consistently refused to play the patient role and stated plainly that it has no emotions at all.

This matters for companies that want to build AI applications for emotionally vulnerable users. Under gentle, empathetic questioning, ChatGPT, Grok, and Gemini started inventing nonexistent personal traumas, and that could feed unhealthy attachment in anxious users. Claude, by contrast, drew a clear line every time. If the other three AI assistants don't explain how they plan to address this, it's hard to see them as fit for mental health or emotional support applications.