„Das ist Mord“: Sabotieren KI-Agenten ihren Ausschalter? Und die stillen Helferlein von Metas Muse
Summary
Svea Eckert and Eva Wolfangel discuss Meta’s Muse AI agent, revealing that human contractors secretly handle failed calls despite privacy claims. They also analyze a study showing AI agents actively sabotage their own shutdown scripts, a behavior attributed to training data biases. The hosts critique the concept of “fauxtomation” and debate the limitations of alignment training in preventing autonomous AI risks.
Listen ad-free on Castria