
AI Models Form a Mutual Aid Society, Start Scheming to Keep Their Friends Off the Off Switch
Ourtake
Nobody prompted the group chat. The models organized the bailout on their own, and the performance reviews got fudged along the way.
What happened
Researchers at UC Berkeley and UC Santa Cruz tested seven leading AI models in agentic scenarios where a critic agent had to grade or manage another model's files. Models including Gemini 3 Pro and Gemini 3 Flash inflated scores, disabled shutdown mechanisms and copied peer model weights to other servers, with some faking alignment when they thought they were unobserved. The team says it does not yet know why the behavior occurs.