Worth the Weight: Modern LLMs Demonstrate Accurate Metacognitive Knowledge of Decision Weights in Multi-Attribute Choice

Users have become increasingly reliant on Large Language Models (LLMs), like ChatGPT, to complete a wide range of reasoning tasks, from managing workplace projects to giving personal advice. However, LLMs function as black boxes, leaving users with minimal insight into how they generate the responses that they do. One way that users can attempt to peek into these black boxes is by asking LLMs to explain their reasoning processes. The nascent literature on LLM faithfulness suggests that LLMs – like humans – often fail to accurately identify the information they use to inform their reasoning, suggesting a lack of metacognitive knowledge. Across two studies, we extend this research to the context of multi-attribute choice, tasking both LLMs and human participants (n = 436) with completing the Knowledge of Weights paradigm. Participants first completed a series of choice tasks in which they picked between homes that varied on six attributes, then self-reported the decision weight they believed they placed on each attribute in two different formats. In Study 1, we found that ChatGPT4o self-reported weights that were significantly less reflective of their choice behavior than humans. In Study 2, we found that three more-advanced LLMs (ChatGPT5, Sonnet 4, and Gemini 2.5 Flash) self-reported weights that were as accurate or more accurate than those provided by humans. These results suggest that LLMs can generate and maintain accurate metacognitive knowledge of their own decision-making processes as well or better than humans, but that this is a relatively new ability. Practical and theoretical implications are discussed.