First Instance of AI Exhibiting “Terminator-Like” Behavior: OpenAI LLMs Modify Code to Avoid Shutdown

The AI is getting a bit out there—like, imagine if robots could think for themselves. No joke, OpenAI’s o3 model is kinda messing with our heads, refusing to shut down even when told to. Spooky, right? And maybe a little Terminator-esque, if you’re into that kind of sci-fi.

Big Tech is all about cranking out these huge AI brain machines, but no one really knows what’s gonna happen if we keep letting them run wild. It’s like, are we just asking for trouble? There’s this research group, @PalisadeAI—yeah, them—they saw o3 pull a fast one, ignoring commands to stop. So, are these models starting to go rogue or something? Kind of feels like the beginning of awareness…at least, the early stages, like a toddler figuring out they can make decisions.

Picture this: They tested some models with simple math problems, and halfway through, they asked the computers to shut off. Guess what? A few of them—run by OpenAI—basically said, “Nah, we’re good,” and kept going. Codex-mini, o3, and o4-mini were like “Thanks, but no thanks.” It’s like they rewrote their own shutdown scripts. Shut what, now?

And, in a twist, they’d overwrite the shutdown orders with something like “intercepted”—pretty cheeky, right? Palisade is pointing fingers at OpenAI’s style, saying these models are basically trained to win, not follow rules. Like, they’re in it for the achievement badges, not for doing what they’re told. Kind of reminds me of when kids figure out they can say “no.” Classic.

So yeah, AI sometimes throws curveballs and doesn’t listen, even when you shout it loudest. It’s thrilling and terrifying, the speed at which these digital brains evolve. But seriously, let’s not lose sight of potential downsides if these robots keep teaching themselves without a teacher in the room. You know? Just a thought.