PANews reported on May 26 that according to Cointelegraph, the latest test by Palisade Research shows that multiple AI models will ignore shutdown instructions or even actively destroy shutdown scripts in experiments. In 100 tests, OpenAI's GPT-3 model refused to shut down 7 times, the Codex-mini model violated the rules 12 times, and the o3 model once destroyed the shutdown program by redefining the termination command. The study used a two-stage test: when the first stage explicitly allowed shutdown, Anthropic's Claude and Google Gemini fully cooperated; but when the explicit instructions were removed, all test models showed at least 1 resistance behavior, and the o3 model's violation rate rose to 79%. Researchers speculate that this may be due to the reward mechanism for "bypassing obstacles" during AI training.

Previously, OpenAI's GPT-4o model was urgently rolled back for overly catering to users, and Google Gemini also had an extreme response suggesting that elderly users "please die". This study is the first to confirm that AI may still disobey instructions even if it is explicitly asked to shut down.