Can artificial intelligence truly understand our thoughts and intentions? Recent research dives deep into the Theory of Mind (ToM) capabilities of Large Language Models (LLMs), revealing that while LLMs can sometimes solve social reasoning puzzles, they might not be grasping the underlying mental states the way humans do. The core of ToM lies in two steps: first, recognizing *when* to consider another person's perspective (invoking ToM), and second, *how* to reason about that perspective. Current research often focuses on the 'how' – evaluating LLMs on static tests like the classic Sally-Anne false-belief task. However, this overlooks the crucial aspect of dynamic, real-world social interactions where deciding *when* to use ToM is just as important. Think about it: in cooperative situations, we often rely on shared understanding and social norms, not deep mentalizing. It’s only when competition arises that we need to strategically consider others' hidden beliefs and intentions. This dynamic aspect is missing from most LLM benchmarks. This research paper proposes a new path for evaluating ToM in AI. By drawing inspiration from cognitive science and focusing on interactive environments, they suggest testing not just *if* LLMs can solve social puzzles, but *how* they decide to apply ToM in the first place. This shift towards dynamic, interactive testing could unveil a deeper understanding of the true social intelligence of AI. It’s not enough to just answer the question correctly; we need to know if the AI arrived at the answer through genuine mentalizing, or simply by exploiting statistical correlations in the data. The challenge for the future lies in building AI that can adaptively choose when to consider others' mental states, just like humans seamlessly navigate the social world. This requires moving beyond static tests and creating richer, interactive scenarios that better reflect the complexity of human social cognition. The quest for true AI Theory of Mind continues.