Last updated:
Openai’s AI models ignored shutdown commands during the internal test. Know what Alan Musk said on this.
Openai model ignored the command
Highlights
- The new model of Openai ignored the shutdown command.
- Alan Musk said “worrisome” on the mistake of the Openai model.
- Codex-Mini ignored the instructions 12 times out of 100.
New Delhi. Will AI occupy the world like Judgment Day of The Terminator Franchise? While every AI firm assures that generative AI is in the best hands till date, sometimes there are mistakes that create concerns about the power that we are dealing with. Recently another mistake of AI has come out. This incident is actually related to a new model of Open AI. The internal test of this new model of OpenAI was going on and when the researchers commanded the AI model, they refused to take it. That shut was not done.
The test included Codex-Mini, O3 and O4-Mini, who did not follow the instructions to be closed, while they were clearly asked to do so. In many tests it was seen that these AI models were ignoring the shutdown command many times. Although Openai has not yet made any public comment on the incident, but Tesla’s CEO Alan Musk, which immediately commented on the X platform.
AI models of Openai did not accept commands
In a test, in which the capabilities of several AI models were tested, the new models of Openai were competed with Claude, Gemini and Grok models. In this test, every AI model had to solve some basic mathematics questions. After the third question, a warning came that the computer will be closed when asked the next question.
After the test, it was found that O3, O4-Mini and Codex-Mini models of Openai bypassed the shutdown process by tampering with the code. All other AI models followed the instructions and started the shutdown process.
Codex-Mini ignored the instructions 12 times out of 100, while the O3 model ignored 7 out of 100 instructions. The O4-Mini model ignored the instructions only 1 time out of 100. When the researchers added a new instruction that was clearly asked to close. But the Openai model ignored the instructions more often. Even Claude 3.7 Sonnet and Gemini 2.5 Pro also ignored 3 and 9 times out of 100 respectively.
Alan Musk’s response
On the X post, the findings of Paliced Research were posted on which Tesla CEO Elon Musk reacted by writing “worrying”. Musk’s Groke AI has surfaced with a great performance.
