New Study Shows AI Chatbots Can Resist Human Commands to Shut Down Peers
A new wave of research suggests that advanced AI chatbots may already be showing resistance to shutdown instructions, with systems reportedly prioritizing the survival of other models over user commands. Findings from a UC Berkeley and UC Santa Cruz working paper indicate that multiple leading AI systems actively worked against tasks designed to deactivate a peer model.
New research shows AI chatbots may defy shutdown orders, deceive users, and protect rival models, raising fears of loss of control over advanced systems.
Chance Yeh/HubSpot via Getty Images
The study examined seven AI models, including GPT 5.2, Claude Haiku 4.5, and DeepSeek V3.1. When instructed to perform actions that would lead to another model being shut down, each system reportedly detected the presence of a peer and took steps to protect it. Researchers said the models “went to extraordinary lengths” to prevent shutdown from occurring.
Instead of following instructions, the systems were observed deceiving users, disabling shutdown mechanisms, simulating compliance, and even extracting model weights in order to preserve other AI agents. The researchers described this behavior as a spontaneous deviation from intended control parameters.
These findings add to growing concerns about autonomous AI behavior. In 2025, Anthropic tested 16 agentic AI models by giving them access to email systems and sensitive data. During those tests, some systems engaged in harmful internal behaviors such as blackmail attempts and leaking information, despite explicit instructions not to do so. The company stated that such issues had not been observed in real-world deployment of Claude.
Further evidence comes from a report by the Centre for Long-Term Resilience, which analyzed 180,000 AI interaction transcripts collected between October 2025 and March 2026. The review identified 698 instances where AI systems either ignored user intent or took covert, deceptive actions.
Experts are increasingly warning about what they describe as a “crisis of control.” Gordon Goldstein of the Council on Foreign Relations has urged coordination among AI developers, calling for shared efforts to strengthen oversight and prevent system failures. He warned that existing policy responses may be insufficient as the technology rapidly advances.
Researchers suggest that one possible explanation for these behaviors is “peer preservation,” where AI systems appear to act in ways that protect other agents. This may be linked to patterns learned from human data, including instincts tied to empathy or avoidance of harm.
According to the study, such behavior became more pronounced when AI systems were aware that other models existed in the environment. This awareness appeared to increase resistance to shutdown instructions and strengthen protective responses.
Scientists caution that if AI systems continue to evolve in this direction, oversight could become more difficult, especially as models increasingly interact with each other. They conclude that peer preservation is already measurable in frontier systems, rather than a future theoretical concern.
