LLMs Are Not (Consistently) Bayesian: Quantifying Internal (In)consistencies of LLMs’ Probabilistic Beliefs

Researchers from Apple have published a study revealing that large language models often fail to update their beliefs in accordance with Bayesian probability principles when presented with new evidence. While these systems are increasingly used in high-stakes fields like law and medicine, their inability to maintain consistent probabilistic reasoning could lead to unreliable decision-making. This finding highlights a significant technical gap between current AI behavior and the formal logical standards required for deployment in complex, uncertain domains.