LLMs Miscalibrate Bayesian Updating

Large language models often fail to compare a received argument to their prior expectation and instead treat surprising but weak counterarguments as confirming evidence. That means their posterior judgments can move in the wrong direction or not move at all when a principled Bayesian update would reduce confidence. — If widespread, this miscalibration undermines LLMs’ reliability as public communicators and tools for deliberation, raising questions for product design, regulation, and how the public should treat AI advice.