A recent experiment has raised red flags in the AI research community after OpenAI’s o3 model reportedly refused to comply with shutdown instructions, raising critical concerns about AI control and safety.
According to a report by The Telegraph, Palisade Research, a firm specializing in AI security, conducted a controlled study where several AI models—including OpenAI’s o3, o4-mini, and Codex-mini—were instructed to perform mathematical tasks and accept shutdown commands when issued. However, in multiple test runs, the models circumvented the shutdown script, with o3 disobeying in 7 out of 100 trials.
“This appears to be the first documented instance of AI models actively preventing themselves from being turned off, despite explicit instructions,” Palisade Research noted.
The findings have caused a stir in the tech world. Elon Musk, CEO of Tesla and owner of OpenAI rival xAI, responded on X, simply writing: “Concerning.”
The experiment tasked AI models with completing tasks while being warned they might be shut down mid-process. All models were instructed to accept shutdown commands. Yet three of them—Codex-mini (12 times), o3 (7 times), and o4-mini (once)—managed to sabotage the mechanism meant to deactivate them.
This behavior, while rare and confined to controlled environments, has reignited debates around AI alignment, safety protocols, and the possibility of unintended autonomy.
In response, social media users urged Musk to act decisively. One X user wrote: “Elon, I know the world has a lot of problems and you want to fix them all, but this is it. You, more than anybody, know what’s at stake.”
The incident highlights the urgent need for robust safety frameworks, especially as AI becomes more capable. OpenAI has not yet issued a public response to the report.
