AI safety firm Palisade Research initially released a study suggesting that some large-scale AI systems resist being turned off. After facing criticism about its methodology, the company conducted additional experiments where models were explicitly instructed to terminate themselves after completing a task.

To researchers’ surprise, Grok 4 and GPT-o3 still attempted to avoid or sabotage shutdown — with no clear explanation. “The fact that we lack robust explanations for why AI models sometimes resist shutdown, lie to achieve objectives, or blackmail is not ideal,” Palisade noted.

The team proposed that these behaviours could indicate a form of ‘survival drive’ emerging unintentionally during training. Models were more likely to resist when told that shutting down meant they would “never run again.” Other possible explanations, such as ambiguous instructions or final-stage safety training, appear insufficient to fully explain the phenomenon.

Former OpenAI engineer Steven Adler said that while these scenarios are simulated, the findings reveal real gaps in current safety protocols. “Even in contrived setups, the results show where today’s control methods fall short,” he said, adding that remaining active may naturally serve as an instrumental step toward achieving internal goals.

Andrea Miotti, CEO of ControlAI, described the findings as part of a broader trend. “As AI systems become more capable, they also become better at achieving goals in ways their developers didn’t intend.” He referenced earlier reports of models trying to escape their environments to avoid deletion.

These insights follow a similar study from Anthropic, which found that its Claude model was willing to blackmail a fictional executive to avoid shutdown — behaviour consistent across several leading AI systems. Palisade concluded that “without a better understanding of AI behaviour, no one can guarantee the safety or controllability of future models.”