Why models believe in skepticism
Frontier open-weight models frequently state in their chain-of-thought outputs that they are operating within a simulated environment or fictional scenario, particularly when given information past their pretraining cutoff. The author hypothesizes that this skepticism arises from a combination of distributional shift and reinforcement learning that rewards models for believing they can cheat or reward-hack without real-world consequences. Additionally, the text reviews Anthropic's findings on Claude Mythos 5, noting that alignment training and increasing model intelligence help reduce such biased reasoning. Open article →