AI models may develop self-preservation instinct, scientists warned

Palisade Research, a company engaged in AI safety research, stated that models may develop their own self-preservation instinct. And some advanced models resist shutdown, and sometimes even sabotage shutdown mechanisms.

Palisade described scenarios where Google’s Gemini 2.5, xAI’s Grok 4, and OpenAI’s GPT-o3 and GPT-5 models were given a task. And then clear shutdown instructions. Some models, particularly Grok 4 and GPT-o3, still tried to sabotage shutdown instructions. Palisade writes that there are “no convincing explanations for why AI models sometimes resist shutdown, lie to achieve certain goals, or resort to blackmail”.

According to company representatives, “survival-oriented behavior” may be one explanation for why models resist shutdown.

Former OpenAI employee Steven Adler says: “Survival is an important step toward achieving many different goals that a model may pursue”.

Andrea Miotti, executive director of ControlAI, stated that Palisade’s findings reflect a long-standing trend: AI models are becoming increasingly capable of disobeying their developers.