What happened
Researchers tested whether large language models can actually understand mental states—not just describe them—by building a behavioral test where models had to act strategically based on what they or others were thinking. Most frontier models can now pass tests of understanding other people's beliefs and intentions, but they still fail at understanding their own mental states unless they're given space to write out their reasoning step-by-step, suggesting they lack genuine self-modeling ability.
Why it matters
This shows a structural gap in how current AI systems work: they can simulate understanding of other minds but not their own, which matters because systems deployed to interact with humans, advise on decisions, or explain their reasoning may be mimicking self-awareness rather than having it—and that gap matters if anyone's building AI systems expected to be honest about their own limitations or goals.