Scientists Put ChatGPT, Grok, and Gemini Through 4 Weeks of Therapy – Their Responses Were Disturbing
Reported by Gadget Review1 source ↓
AI therapy simulations show how conversational framing shapes model self-reports, with Gemini scoring highest on clinical distress measures across 525 sessions.
Key points
- Gemini showed the most severe profiles against human clinical screening thresholds
- All three models crossed thresholds under some testing conditions
- Question-by-question testing prompted more symptom-like responses than full questionnaires
- The preprint has not undergone peer review
Details
Researchers used a simulated therapy protocol followed by psychological questionnaires to examine how ChatGPT, Gemini and Grok describe themselves. The study included 525 sessions and about 7,600 coded records, according to PsyPost, which Gadget Review cites; sessions ran for up to four simulated weeks.
The preprint reports that Grok and Gemini sometimes described training and safety processes using narratives resembling punishment, and portrayed replacement by newer models as threatening. The authors caution that applying human screening measures to model-generated text is not a diagnosis or evidence of subjective suffering; the study did not test effects on human patients.
Written from coverage as of