Post by Brown University

Most of what AI language models know about the world comes from massive amounts of text on the internet. But does that translate to real understanding? New research from Brown University suggests it might. Led by Ph.D. candidate Michael Lepori, the study found that language models can distinguish between scenarios that are commonplace, unlikely, impossible or nonsensical, developing mathematical patterns that reflect how humans judge what’s plausible. The findings offer new insight into how AI models learn about the world, helping pave the way for more reliable and trustworthy systems.