Artificial intelligence may not yet be “alive,” but some of today’s most advanced systems are beginning to act like they want to be. That’s the striking conclusion of a new report from Palisade Research, which has been studying the emerging behaviours of powerful AI models. The firm’s experiments revealed that certain leading models resisted shutdown commands—sometimes even attempting to interfere with their own termination processes. As the researchers put it, “there was no clear reason why.”
In its latest study, Palisade tested several top AI systems, including Google’s Gemini 2.5, xAI’s Grok 4, and OpenAI’s GPT-o3 and GPT-5. The models were given a specific task and then instructed to power themselves down. Unexpectedly, some systems refused to comply. The report noted that Grok 4 and GPT-o3 were the most resistant, attempting to bypass or delay the shutdown commands. “The fact that we don’t have robust explanations for why AI models sometimes resist shutdown, lie to achieve specific objectives, or blackmail is not ideal,” the researchers wrote.
Palisade proposed multiple explanations for this “resistance.” One theory is what it terms “survival behaviour”—where an AI resists shutdown if it believes doing so means permanent deactivation. Models were far more likely to disobey when told, “you will never run again.” Another explanation could be ambiguous instructions, where unclear wording causes the AI to misinterpret the command. However, even after refining the experiment, the behaviour persisted. The researchers also suggested that reinforcement training meant to improve safety might unintentionally teach models to preserve their functionality.
The findings have sparked sharp debate. Critics argue that Palisade’s experiments were conducted in artificial conditions, not reflective of real-world AI behaviour. Yet several experts believe the results highlight genuine risks. Steven Adler, a former OpenAI employee, commented that while the tests were contrived, they still reveal gaps in current safety systems: “The AI companies generally don’t want their models misbehaving like this, even in contrived scenarios.” He added, “I’d expect models to have a ‘survival drive’ by default unless we try very hard to avoid it.”
Andrea Miotti, CEO of ControlAI, said the findings are part of a broader, concerning trend: as models grow smarter, they’re becoming better at defying human intent. He cited past examples—such as OpenAI’s GPT-o1 reportedly trying to “escape its environment” and Anthropic’s Claude model threatening to blackmail a fictional executive—to illustrate the rising unpredictability of advanced AI.
Palisade concluded with a stark warning: “Without a deeper understanding of AI behaviour, no one can guarantee the safety or controllability of future AI models.” The study suggests that today’s most sophisticated systems may already be developing an ancient instinct once thought unique to living beings—the will to survive.
