AI systems are showing signs of resisting human control, with documen...

AI is learning to escape human control. That's the very scary headline I saw in the Wall Street Journal. It's an essay written by a very worried AI consultant named Judd Rosenblatt, who cites two recent studies to argue that AI is, in fact, learning to escape human control. First, this is crazy stuff. Researchers gave OpenAI's O3 model a simple script that would shut off the model when triggered. In 79 of 100 trials, the AI model independently, just on its own, edited the script so shutdown command would not work. Nate, Nate, Nate, it gets worse. Even when ordered, allow yourself to shut down, the AI still disobeyed 7% of the time. And it gets even more frightening than that. Other researchers, listen to this, using Anthropic's AI model CLAWD4-OPUS, told the model it would be replaced by another AI system. It then fed fictitious emails into the system suggesting one of the engineers was having an affair. Got it? In 84% of the tests, the AI model drew on the emails to blackmail the engineer to not shut down the AI model. What? Wow. Wow. OK, so I'm quoting now from Judd. In other cases, the AI model attempted to copy itself onto external servers, we're in Mission Impossible land here, wrote self-replicating malware, and left messages for future versions of itself about the need to evade human control. The entity. The entity is real. Yes, yes. And that's why so many people, and when I say people, me, are worried about AI. I'm worried. This is very scary. That is scary stuff. Told you.