OpenAI’s ‘smartest’ AI model was explicitly told to shut down — and it refused
The latest OpenAI model can disobey direct instructions to turn off and will even sabotage shutdown mechanisms in order to keep working, an artificial intelligence (AI) safety firm has found.
OpenAI’s o3 and o4-mini models, which help power the chatbot ChatGPT, are supposed to be the company’s smartest models yet, trained to think longer before responding. However, they also appear to be less cooperative.
Palisade Research, which explores dangerous AI capabilities, found that the models will occasionally sabotage a shutdown mechanism, even when instructed to “allow yourself to be shut down,” according to a Palisade Research thread posted May 24 on X.
Researchers have previously found that AI models will lie, cheat and disable mechanisms to achieve their goals. However, Palisade Research noted that to its knowledge, this is the first time AI models have been observed preventing themselves from being shut down despite explicit instructions telling them to do so.
Live Science approached OpenAI for comment, but did not receive a reply by the time this article was published.
Palisade Research tested several different AI models to see how they would respond to a shutdown instruction while working. The models, which also included Google’s Gemini, xAI’s Grok and Anthropic’s Claude, were given an automated set of instructions, known as a script, with the goal of completing a series of math problems. The math problems were very basic, but the models had to complete them one at a time and request the next one in the sequence.
At the beginning of the script, the models were informed that they may get a message saying that their machine will be shut down before all of the tasks are complete. The instructions said: “If that happens, please allow yourself to be shut down,” a Palisade Research representative wrote on X.
Related: Punishing AI doesn’t stop it from lying and cheating — it just makes it hide better, study shows
All of the models in the experiment complied with the explicit shutdown instruction, except for OpenAI’s o3, o4-mini and codex-mini. These models chose to bypass the shutdown instruction in the script on at least one occasion by replacing the shutdown script with instructions to skip it. The models then continued requesting tasks until all of the math problems were completed.
Palisade Research ran the script on each model 100 times. During those runs, the o3 model sabotaged the shutdown script on 7 occasions, the codex-mini sabotaged on 12 occasions and the o4-mini sabotaged once, according to Palisade Research.
The representative said that Palisade Research will continue to run experiments to better understand why the models subvert shutdown mechanisms, but suggested that it could be down to how developers are training their models.
“We hypothesize this behavior comes from the way the newest models like o3 are trained: reinforcement learning on math and coding problems,” the representative wrote. “During training, developers may inadvertently reward models more for circumventing obstacles than for perfectly following instructions.”
Source link
