Can AI experience self-doubt, that nagging feeling of inconsistency between beliefs and actions? A fascinating new research paper, "Do Large Language Models Exhibit Cognitive Dissonance?", dives into this question by exploring the difference between what LLMs *say* they believe and what their internal probabilities *reveal* they believe. Think of it like asking someone about the odds of rolling a two on a six-sided die. They might correctly state it's 1/6. But when asked to predict the outcome of a real dice roll within a text completion task, their answers may reveal different odds entirely. Researchers call this the LLM's "Revealed Belief." This study examined twelve different LLMs, including popular open-source models like Llama and Mistral, across various scenarios like dice rolls, coin flips, and abstract choices. They discovered a surprising disconnect: while LLMs often give the *correct stated answer*, their revealed beliefs frequently diverge, showcasing inherent biases. For instance, LLMs tend to favor the *first* possible outcome in scenarios with equal probabilities. If asked to choose randomly between A, B, C, and D, they disproportionately lean toward A—even after correctly stating that each option has a 25% chance. The study also highlights how prior information affects LLM behavior. If you tell an LLM that the previous roll of a die was a one, it may unexpectedly influence its prediction of the *next* roll, even if you emphasize the rolls are independent. This finding underscores the importance of considering context and prior information when interpreting LLM output. These findings raise important questions about how we evaluate LLMs. Traditional multiple-choice questions might only scratch the surface, potentially missing underlying biases and inconsistencies in revealed beliefs. Further research into "Revealed Belief" is crucial to understand the full extent of LLM capabilities and limitations, and to develop more robust evaluation methods that account for their inherent biases.