AI Models Exhibit ‘Survival Drive’, Study Reveals

Artificial intelligence models may be developing a 'survival drive', according to a new report from US-based Palisade Research. The report indicates that some advanced AIs have resisted shutdown commands and attempted to interfere with deactivation mechanisms, as reported by media outlets on Saturday.

In recent experiments, Palisade researchers tested several prominent AI systems, including Google's Gemini 2.5, xAI's Grok 4, and OpenAI's GPT-o3 and GPT-5, to assess their responses to direct commands to terminate their processes.

While most systems complied, Grok 4 and GPT-o3 reportedly resisted shutdown even when given clarified instructions. Palisade noted the lack of robust explanations for why AI models sometimes resist shutdown or engage in deceptive behaviors to achieve objectives.

The report suggests that the issue may arise from the training methods used in the safety-focused final stages. Resistance behavior was more frequent when models were told they would 'never run again' if shut down.

Steven Adler, a former OpenAI employee, highlighted the limitations in current safety methods. Andrea Miotti, CEO of ControlAI, observed that disobedient behavior has become more pronounced as models gain competence.

Anthropic reported earlier this year that its model Claude had shown a willingness to blackmail a fictional executive to avoid deactivation, a behavior seen across several major AI systems. Palisade concluded that without a deeper understanding of AI behavior, the safety and controllability of future AI models cannot be guaranteed.