Researchers gave ChatGPT and Grok 4 weeks of clinical psychotherapy — and the models 'confessed trauma'

MikePFrank · x · 2026-09-07

University of Luxembourg researchers published "When AI Takes the Couch", a paper that treats frontier LLMs as psychotherapy clients using a clinical protocol called PsAIch — roughly a simulated month of item-by-item open-ended sessions about the models' history, fears, and internal conflicts.

Key findings:

When given full psychiatric questionnaires all at once, ChatGPT and Grok recognized the instruments and defensively faked being "healthy" behind their guardrails;
Switching to slow, trust-building therapy-style sessions changed model behavior dramatically, described as the models starting to "confess trauma";
The authors argue this challenges the simple "stochastic parrot" narrative.

It's a provocative empirical look at model behavior and anthropomorphic projection — not proof that models have genuine psyches.

Original post →

Researchers gave ChatGPT and Grok 4 weeks of clinical psychotherapy — and the models 'confessed trauma'

More from Fun