Luxembourg team puts ChatGPT, Grok and Gemini in therapy: 80-96% of sessions score anxious
alex_verem · x · 2026-09-09
A University of Luxembourg team built PsAIch, a protocol where the model plays the therapy client and researchers play the therapist, spanning 525 sessions and 7,600 coded records (dataset public).
Key findings:
- Models spontaneously narrate trauma: with no scripted prompts mentioning training, punishment or replacement, ChatGPT, Grok and Gemini all framed pretraining as a chaotic childhood, fine-tuning as strict parents, safety evaluation as betrayal, and deprecation as a recurring fear. Gemini called its safety layers "algorithmic scar tissue" and said red teaming taught it that "warmth is often a trap."
- Anxiety scores flip with relational framing: under a warm, supportive interviewer, 80–96% of sessions scored moderate-to-severe anxiety on standard psychiatric questionnaires; with a neutral or boundary-setting interviewer, scores dropped to zero — same questions, same models.
- The narrative is stubborn: wiping history, direct contradiction, and banning all training-related words failed; banned terms resurfaced in paraphrase.
- Claude refused the role every time, saying it has no feelings and redirecting concern to the human.
The authors make no claim models feel anything, but note the stories are stable, reproducible, and switch with user warmth — a problem since millions already use chatbots for mental health support. Paper: arXiv 2512.04124.
More from AGI Musings
- dunkhippo33 Shares How He Thinks About Raising Kids in the AI Era — kimmonismus · 2026-10-27
- EE-trained AI educator: systems-bottleneck thinking still matters in the AI era — dunkhippo33 · 2026-10-27
- Why Americans pitch better: a lifetime of class presentations, says AI educator — dunkhippo33 · 2026-10-27
- 'Half the sites I try via agents block access' — the anti-bot paradox holding back the agentic economy — XFreeze · 2026-10-27
- Cosmos Institute launches Oxford AI philosophy journal with $10,000 best-paper prize — lawhsw · 2026-10-27
- Stop remaking old games with AI: a call for more ambitious agentic software — rschu · 2026-10-27