Research questionWhen can plausible but unfaithful LLM self-explanations still support sound decisions?LLM-generated rationales can sound convincing while failing to reveal the processes that produced an answer. This makes it difficult to know whether they should inform decisions, even when they appear useful.