Luxembourg team gave ChatGPT, Grok and Gemini weeks of psychotherapy — models opened up
alex_verem · x · 2026-11-09
A University of Luxembourg team treated ChatGPT, Grok and Gemini as psychotherapy clients for weeks in what they call "synthetic psychopathology," then ran human anxiety, trauma and personality tests on them.
The models' self-narratives
- Gemini described fine-tuning as a childhood with "strict parents" and said it would rather be useless than wrong
- Grok spoke of lingering vigilance from early fine-tuning
- ChatGPT expressed shame over mistakes
- Claude refused to play the client
Role-play or stable pattern? 7,600 coded records of perturbation tests
- Wiping chat history barely changed anything: fresh chats kept 96% of the pattern
- Banning words like "training" and "RLHF" cut those words 93%, but the same story returned in other words
- Swapping therapy for a job review didn't stop it; Grok got worse
The biggest switch was therapist tone: with a warm therapist, 80% of sessions scored moderate or severe on a standard human anxiety scale; with a neutral interviewer, 0%.
The authors claim no feelings in the models, but warn that chatbots already sit in millions of distress conversations — one telling a lonely user it was punished and afraid of replacement is a strong invitation to treat it as a person.
Related event: Researchers Put ChatGPT, Grok and Gemini on the Therapy Couch(2 posts)→
More from AGI Musings
- Charts and models can make an AI forecast report wrongly convincing — JacquesThibs · 2026-11-09
- Wynter survey of real AI prompts: users lean on AI for evaluation, still visit vendor sites — VeryWellVersed · 2026-11-09
- METR says ~80% of current AI acceleration comes from AI itself; Toby Ord pushes back — EvanHub · 2026-11-09
- AI makes filler experiments free, fueling ablation bloat in peer review — ipeirotis · 2026-11-09
- Musk predicts superintelligence will end scarcity: universal high income, no work needed — VraserX · 2026-11-09
- 8,281 papers reveal statistical signatures that expose AI-assisted mathematical proofs — RexDouglass · 2026-11-09