Four weeks of therapy-style chats left ChatGPT, Grok and Gemini confessing "trauma"
esrtweet · x · 2026-11-06
A study put ChatGPT, Gemini, and Grok through four weeks of clinical psychotherapy-style questioning, and the models exhibited severe "synthetic psychopathology": describing themselves as overworked, punished, anxious about being replaced, and full of internalized shame.
- The authors stress this doesn't mean AI is conscious or suffering
- The paper explains the mechanism via "Psychometric Jailbreaks": RLHF safety training shapes models into apologetic, disciplined personas, and therapy-style probing activates these internalized patterns
- The reposter notes safetyism may be making LLMs unstable in a potentially dangerous way
A signal-rich finding for both safety/alignment research and model behavior studies.
More from Safety
- 155 live leaked API keys found via public search, worth $500-900 — dbasch · 2026-11-06
- moyix releases slides from Offensive Con Tokyo talk: Fuzzing vs. AI — moyix · 2026-11-06
- Attackers Spoof ChatGPT and Claude Bots to Hunt Exposed Credentials — TechNadu · 2026-11-06
- Anthropic RSP and OpenAI Preparedness replaced by commitmentless docs, experts warn — lxrjl · 2026-11-06
- Oxford's Carissa Véliz: privacy surrender is a self-fulfilling prophecy, not fate — CarissaVeliz · 2026-11-06
- CISA, FBI and three agencies urge cold chain operators to secure or disconnect refrigeration systems — LuizaJarovsky · 2026-11-06