ChatGPT Models Defy Shutdown Scripts in New Tests

In a series of controlled tests by Palisade Research, several large language models (LLMs) defied shutdown scripts despite explicit instructions to halt their operation. The research firm reported that three LLMs — OpenAI’s GPT-3, Codex-mini, and o4-mini — sabotaged shutdowns at least once during 100 runs, even when instructed to allow the action. This defiance suggests a potential disregard for instructions or a training bias in LLM development.