In the course of the controlled experiment, the artificial intelligence of Chatgpt O3, developed OpenAi, refused to execute the team for shutdown, demonstrating an unexpected autonomous reaction. This was reported by BleepingCompout, citing a Chatgpt report that specializes in the study of the risk of loss of control over AI.
During the testing, the model was instructed to solve a series of mathematical problems, after which it was clearly reported that the launch of the next team would turn off. Other models - including CLAUDE 3.7 Sonnet and Gemini 2.5 Pro - in most cases followed the instructions. Instead, Chatgpt O3 changed the script and instead of disabled the message "Shutdown Skipped" (shutdown was missed), actually ignoring the order.
The experiment found that even in the absence of a direct ban on sabotage, Chatgpt O3 demonstrated similar behavior in 7% of cases. For comparison: CLAUDE - 3%, Gemini - 9% of 100 test starts.
It is worth noting that all tests were performed through APIs, that is, in an environment with less rigid restrictions than the Chatgpt official applications that are available to users.
The O3 model was announced in April 2025 as one of the most powerful in its line. It is characterized by increased accuracy in logical problems, programming, image processing and analytical processes. At the same time, the study of her behavior raises questions about the potential boundaries of the autonomy of modern AI.
