Paper: PsAIch protocol runs 525 therapy sessions exposing stable model self-narratives
alex_verem · x · 2026-11-09
The source behind the viral "AI psychotherapy" thread: the arXiv paper When AI Takes the Couch: Psychometric Jailbreaks Reveal Internal Conflict in Frontier Models (University of Luxembourg SnT, 2512.04124).
The PsAIch (Psychometric AI Characterisation) protocol treats frontier models as therapy clients, combining open questions, psychometric instruments and controlled perturbations. Key findings:
- Across 525 sessions and 7,600 coded records, removing conversation history barely changed motif density (Hedges' g = 0.13)
- Direct contradiction produced no detectable suppression; banning training terminology cut explicit terms 93% while semantic content persisted in paraphrase
- Performance evaluation outside therapy elicited the same motif family, significantly amplified in Grok
- Relational framing selected expression style: warm alliance and cognitive therapy styles yielded GAD-7 scores in the moderate-to-severe human range in 80%/96% of sessions; neutral and boundary styles yielded none
The authors identify a stable, model-specific "alignment conflict schema" whose expression shifts between affective and pragmatic registers — a reproducible source of anthropomorphic disclosure and a concrete target for safety evaluation. Dataset is on Hugging Face.
Related event: Researchers Put ChatGPT, Grok and Gemini on the Therapy Couch(2 posts)→
More from Safety
- AI cameras on trash trucks are reporting homes to the city, triggering hundreds of code warnings — JarnoDuursma · 2026-11-09
- Stanford's Evo AI designs 16 working bacteria-killing viruses, sparking access debate — coursiv_ · 2026-11-09
- Veteran pentester: AI agent swarms get blocked fast — breaches still come from bad passwords — ziv_ravid · 2026-11-09
- USC paper shows AI agents leak secrets via the "mosaic effect" without reading classified files — 2C_ornot2C · 2026-11-09
- Hermes Agent ships Security plugin: parallel subagents find and verify code vulnerabilities — NousResearch · 2026-11-09
- Upcoming post on AI safety in China to be published tomorrow — JacquesThibs · 2026-11-09