Imagine an AI facing a classic philosophical puzzle: Newcomb's Paradox. It has to choose between two boxes: one transparent with a visible reward, the other opaque, potentially containing a much larger prize or nothing at all. A super-intelligent predictor has already placed the larger prize in the opaque box *if and only if* it predicted the AI would only choose that box. Which should it choose? This dilemma, where predicting the future impacts the present, lies at the heart of a new research paper exploring how AI understands decision theory. Researchers at Carnegie Mellon University and Anthropic have created a dataset of hundreds of complex, Newcomb-like problems, posed as multiple-choice questions. They tested a range of language models, including those from OpenAI, Google, and Anthropic, to see if they could grasp the nuances of decision-making under uncertainty. The results were fascinating. While some top-tier models showed an impressive ability to reason through complex scenarios, even they sometimes stumbled. Intriguingly, the researchers found a strong link between a model's ability to solve these decision problems and its tendency to favor a particular decision-making approach called 'evidential decision theory', which emphasizes choosing actions that correlate with better outcomes. But what does this mean? Are these models actually forming 'opinions' about the best way to make decisions? Or are they simply picking up patterns in the data? This research highlights the challenges of building AI that can reason like humans. The models' struggles with some seemingly simple problems suggest that true, human-like decision-making might require more than just vast amounts of data. Future work will explore how these insights can help build more cooperative and reliable AI, especially in situations where AIs interact with each other—a scenario that increasingly resembles Newcomb’s Paradox itself.