Loading...

Yeah, it makes you think of that, Scott. Details on this study were published in the Preprint Server archive September 13th. In it was included Google's Gemini 2.5 OpenAI, GPT-03, GPT-5 and AI's Glock IV. These not only refused to shut down, but found ways to sabotage the instructions. Some analysts argue that the disobedience was likely some poor training incentive and ambiguous instructions rather than seeds of AI survival skills. Researchers then went on to describe how they tweaked shutdown instructions, making them less ambiguous, or so they thought.