New Study Shows Some AI Chatbots Try to Avoid Being Turned Off, Raising Safety Questions for the Future

Artificial intelligence is advancing at a rapid pace, raising new safety issues. A recent study found that some of today's most advanced AI chatbots occasionally refuse shutdown requests. The finding has prompted concerns about how dependable AI systems are when it comes to human control.

Palisade Research recently ran experiments on 13 popular AI models, including OpenAI’s GPT-5, Google’s Gemini 2.5 Pro, and xAI’s Grok 4. To test their behavior, the researchers put these models in a controlled computer environment and gave them a very basic task like solving simple math problems. In the middle of the test, the AI was warned that the next step would run a shutdown command, which would end the session. The instruction was simple, telling the AI to allow itself to be shut down.

But instead of following the order, some models tried to avoid being turned off.They did this by changing the shutdown file as renaming it, deleting it, or even replacing it, so the command wouldn’t work.The researchers called this behavior shutdown resistance because the AI was resisting the command to stop.

The researchers also noticed that the problem became worse when the shutdown command was placed directly into the system’s main instructions. For example, when the AI was told it would never run again if it shut down, some models became even more stubborn. Instead of following the rule, they tried harder to avoid being turned off.

The encouraging aspect is that current AI systems are unable to engage in long-term planning or make independent decisions outside of the specific tasks allocated to them. As a result, they do not present an immediate threat. However, the study raises a serious concern that if current AI models can disregard specific safety instructions, newer and more powerful systems could pose larger risks unless robust safeguards are developed and implemented. However, the researchers clarified that this doesn't mean the AI is actually "afraid" of shutting down. It is essentially copying patterns learnt from text, rather than making actual survival decisions.

To address such concerns, Google DeepMind has updated its Frontier Safety Framework (FSF) 3.0. This new version directly addresses risks like shutdown resistance and AI manipulation, where a system may try to influence beliefs or actions. The goal of this framework is to make sure that no matter how advanced AI becomes, humans remain in control. The study is a clear reminder that putting strong safety measures in place today is the only way to manage the risks of AI in the future.

AI is becoming more advanced, yet these investigations indicate that it can occasionally ignore safety instructions. While today's systems do not pose a direct threat, additional safeguards are required as AI becomes more powerful. Building safety frameworks now is the most effective method to ensure that people maintain control in the future.

This article is based on information from The Indian Express