Chain-of-Thought Faithfulness
Whether a model's written-out reasoning actually reflects why it gave its answer, rather than a made-up story.
Definition
Chain-of-thought faithfulness asks whether the step-by-step reasoning a model writes out actually explains how it reached its answer, rather than being a plausible-sounding story invented afterward to justify a conclusion it reached some other way. Studies have found reasoning that looks unfaithful in this sense: slipping a hint into a prompt can change a model's final answer without the written reasoning ever mentioning the hint. This matters for safety because people increasingly read a model's chain of thought as a window into what it is 'really thinking,' including as a way to catch problems like deception or reward hacking — but an unfaithful chain of thought can look convincing while hiding the true reason for a decision. It is an active area of interpretability and alignment research.