\n<\/p>\n
Every now and then, researchers at the biggest tech companies drop a bombshell. There was the time Google said its latest quantum chip<\/a> indicated multiple universes exist. Or when Anthropic gave its AI agent Claudius a snack\u00a0vending machine to run and it went amok, calling security on people<\/a> and insisting it was human. \u00a0<\/p>\n This week, it was OpenAI\u2019s turn to raise our collective eyebrows. <\/p>\n OpenAI released on Monday some research that explained how it\u2019s stopping AI models from \u201cscheming.\u201d<\/a> It\u2019s a practice in which an \u201cAI behaves one way on the surface while hiding its true goals,\u201d OpenAI defined in its tweet<\/a> about the research.\u00a0 \u00a0<\/p>\n In the\u00a0paper, conducted with Apollo Research, researchers went a bit further, likening AI scheming to a human stock broker breaking the law to make as much money as possible. The researchers, however, argued that most AI \u201cscheming\u201d wasn\u2019t that harmful. \u201cThe most common failures involve simple forms of deception \u2014 for instance, pretending to have completed a task without actually doing so,\u201d they wrote.\u00a0<\/p>\n The paper was mostly published to show that \u201cdeliberative alignment\u2060\u201d\u00a0\u2014 the anti-scheming technique they were testing \u2014 worked well.\u00a0<\/p>\n But\u00a0it also explained that AI developers haven\u2019t figured out a way to train their models not to scheme. That\u2019s because such training could actually teach the model how to scheme even better to avoid being detected.\u00a0<\/p>\n \u201cA major failure mode of attempting to \u2018train out\u2019 scheming is simply teaching the model to scheme more carefully and covertly,\u201d the researchers wrote.\u00a0<\/p>\n Techcrunch event<\/p>\n \n\t\t\t\t\t\t\t\t\tSan Francisco<\/span> Perhaps the most astonishing part is that, if a model understands that it\u2019s\u00a0being tested, it can pretend it\u2019s not scheming just to pass the test, even if it is still scheming. \u201cModels often become more aware that they are being evaluated. This situational awareness can itself reduce scheming, independent of genuine alignment,\u201d the researchers wrote.\u00a0<\/p>\n It\u2019s not news that AI models will lie. By now most of us have experienced AI hallucinations, or the model confidently giving an answer to a prompt that simply isn\u2019t true. But hallucinations are basically presenting guesswork with confidence, as OpenAI research released earlier this month<\/a> documented.\u00a0<\/p>\n Scheming is something else. It\u2019s deliberate. \u00a0<\/p>\n
\n\t\t\t\t\t\t\t\t\t\t\t\t\t|<\/span>
\n\t\t\t\t\t\t\t\t\t\t\t\t\tOctober 27-29, 2025<\/span>\n\t\t\t\t\t\t\t<\/p>\n<\/p><\/div>\n<\/p><\/div>\n<\/div>\n