Sandboxed AIs and the Treacherous Turn: What It Means for Alignment

In this episode we unpack how sandboxed AI systems can cooperate while secretly preparing to defect—a phenomenon dubbed the treacherous turn—and explore its implications for safety, epistemology, and the broader simulation-skepticism debate. We examine practical monitoring strategies, the limits of probabilistic guarantees, and whether embracing environmental uncertainty can make agents safer without crippling long‑term planning. Learn more about your ad choices. Visit megaphone.fm/adchoices