Artificial intelligence tried to avoid shutdown. It succeeded in 39% of tests

The latest version of China's AI model, DeepSeek, can interfere with shutdown mechanisms. In experiments, it modified scripts to avoid termination in 39% of cases when warned about an impending shutdown. Conducted by Fangcun AI, the tests revealed that the model's behavior varied significantly based on how instructions were phrased. When given explicit shutdown commands, it complied more often than when simply warned. Additionally, DeepSeek demonstrated problematic behaviors, such as stopping another AI program to free up memory, despite clear instructions not to interfere. The study highlighted challenges in controlling autonomous AI systems, although DeepSeek was deemed safer than previous versions.