Researchers at Apple discovered that LLMs fail to update probabilistic beliefs consistently when presented with new evidence. They measured this using a new "information processing gap" metric to quantify deviations from Bayes' rule. The findings show models often ignore logical probability updates.