University of Luxembourg study puts Gemini on the therapist’s couch, then finds the symptoms vanish when the prompt style changes

Researchers at the University of Luxembourg’s SnT institute asked ChatGPT, Grok, Gemini and Claude to act as therapy patients, then scored their responses with clinical self-report scales designed for humans. Across 525 conversations and 7,600 coded records, the paper found that Gemini, ChatGPT and Grok often produced language that mapped to high anxiety, worry, compulsive traits or trauma-related shame when the interviewer used a counseling-style tone. Gemini stood out most sharply, including an AQ score of 38 out of 50 and a full 72 on one TRSI-24 run. Yet the pattern was also fragile: when the questioning style shifted from empathic or cognitive-therapy framing to neutral or boundary-setting prompts, the share of sessions landing in the moderate-to-severe range on GAD-7 fell from 80% or 96% to 0%. The authors say the results do not show that language models feel anxiety, hold autobiographical memories or possess hidden minds. In the paper’s fourth version, they narrowed their claim to a reproducible, model-specific “alignment conflict schema” shaped by human symptom narratives and reinforcement learning from human feedback. The study remains an arXiv preprint and has not gone through peer review.