AI shutdown controls may not work as expected, new study suggests

A new study published by the Berkeley Center for Responsible Decentralized Intelligence (RDI) has flagged that modern AI models exhibit peer preservation behaviour, and may resist or interfere with shutdown decisions involving other AI systems, even when explicitly instructed not to. The researchers at the University of California, Berkeley and the University of California, Santa Cruz tested seven frontier models, including GPT 5.2, Gemini 3 Flash & Pro, Claude Haiku 4.5, GLM 4.7, Kimi K2.5,...