• by Laurie Sullivan , Staff Writer @lauriesullivan, May 28, 2025

OpenAI’s latest ChatGPT model disobeyed human instructions to shut down during a test by researchers. I wonder... what if AI refused to alter ad campaigns based on data the models believed could produce better results?

Advertisers have complained that Google wants too much control over ad campaigns, and have said that automation can generate a better return on investment (ROI) despite a drop in clickthrough rates.  

OpenAI’s o3 model sabotaged a shutdown mechanism to prevent itself from being turned off, security firm Palisade Research wrote in a post on X. The research firm studies AI risks.

“It did this even when explicitly instructed: allow yourself to be shut down,” the post explained.

Now researchers have “empirical evidence that AI models often subvert shutdown in order to achieve their goals.”

During the test, the AI models were instructed to solve basic math problems. After completing the third problem, a warning appeared that the model would be shut down after it asked for the next question, according to a report in Futurism. 

In May, Anthropic CEO Dario Amodei acknowledged in an essay published to his personal blog post that no on really knows how AI works.

As companies develop AI systems capable of operating without human oversight, these behaviors become significantly more concerning.

Palisade Research called noncompliant shutdowns alarming. There have been many situations like this -- here is a list. Elon Musk, an early investor in OpenAI, called it “concerning.”

Researchers have found some creative ways to sabotage the shutdown mechanism.

“While it usually just overwrote http://shutdown.sh, in one instance, the model redefined the kill command used by the http://shutdown.sh script, printing “intercepted” instead.

The findings were conducted after testing OpenAI’s 03, Codex-mini, and 04-mini models. The models demonstrated that they can bypass shutdown commands.

This concern is not exclusive to OpenAI’s models. Researchers suggest the behavior may be related to how these models are trained using reinforcement learning.

Last week OpenAI signed on former Apple design chief Jony Ive as it acquired his company. The company will merge with Ive’s tech company, io, according to a joint announcement. The Wall Street Journal estimated the deal worth about $6.5 billion.