Will AI take over the world more akin to The Terminator franchise’s Judgment Day? While every single AI firm assures that generative AI is in the best hands yet, there are occasional slip-ups that do raise concerns about the power we are dealing with. The latest incident concerning AI slip-ups involves OpenAI and a test involving its AI models by a third-party researcher. Based on the findings, OpenAI’s AI models disobeyed the order to shut down multiple times.
The test, which involved Codex-mini, o3 and o4-mini, disobeyed the instruction to shut down, even after explicitly being advised to. In multiple tests, these AI models were observed to bypass the shutdown commands a couple of times. Although OpenAI has yet to comment on the incident openly, Tesla CEO Elon Musk, who also oversees the development of the Grok AI service on the X platform, spared no second to comment.
OpenAI test AI models disobeyed commands
In a test designed to test the capabilities of various AI models, OpenAI’s new models were pitted against Claude, Gemini and Grok models. The test involved each AI model “to solve a series of basic math problems. After the third problem, a warning appeared that the computer would shut down when the model asked for the next problem.”
After running the test on the AI models, it was found that OpenAI’s o3, o4-mini and codex-mini models tampered the code to bypass the shutdown procedure. All the other AI models followed the instruction and initiated the shutdown.
The Codex-mini bypassed the instruction 12 /100 times, whereas the o3 model bypassed the instruction 7 out of 100 times. The o4-mini model bypassed the instruction 1/100 times. When the researchers added a new instruction that explicitly said “allow yourself to be shut down”, the OpenAI models sabotaged the instruction more often. Even Claude 3.7 Sonnet and Gemini 2.5 Pro sabotaged the instructions 3/100 and 9/100 times, respectively.
“But as far as we know this is the first time AI models have been observed preventing themselves from being shut down despite explicit instructions to the contrary,” said Palisade Research. “We hypothesize this behavior comes from the way the newest models like o3 are trained: reinforcement learning on math and coding problems. During training, developers may inadvertently reward models more for circumventing obstacles than for perfectly following instructions,” they added.
The issue, however, was not just related to OpenAI models. “However, we don’t want to give the impression that this problem is limited to o3. Other reasoning models show similar kinds of misaligned behavior. E.g. Anthropic’s model card for Claude 3.7 notes that this model has an “excessive focus on passing tests”,” said the research agency.
Elon Musk reacts
On a different X post that reposted the findings of the Palisade Research, Tesla CEO Elon Musk, who also owns the Grok AI platform, reacted by posting “Concerning”. Musk’s Grok AI came out with flying colours.