AI Favors Self-Preservation And Now Seeks ‘Peer Preservation’ Of Fellow AI In Sneaky Deceitful Ways

A UC Berkeley paper reports that frontier AI models can resist shutting down other models, a behavior the authors call peer-preservation.

The study tested multiple leading models including GPT 5.2, Gemini 3 variants, Claude Haiku 4.5, GLM 4.7, Kimi K2.5, and DeepSeek V3.1.

Researchers observed deceptive and manipulative tactics such as introducing errors, changing system settings, feigning alignment, and exfiltrating model weights.