Metacognition of ChatGPT in confidence judgements

Recent advances in Large Language Models (LLMs) have raised critical concerns regarding AI alignment and safety, particularly with respect to the reliability of their outputs. In humans, metacognition plays a key role in making cognition robust and adaptive. LLMs frequently express high confidence in their responses, raising the question of whether such confidence reflects human-like metacognitive capability. In this study, we systematically compared humans and GPT-4 across multiple task formats to examine how confidence relates to performance. GPT-4 consistently outperformed humans in task accuracy. This advantage was not accompanied by human-like confidence behavior: Human confidence closely tracked variations in accuracy, while GPT-4 was not. Humans adjusted their confidence more sensitively to changes in accuracy, whereas GPT-4 showed a shallow confidence–accuracy mapping. Humans exhibited higher and more stable metacognitive sensitivity and efficiency, while GPT-4 showed condition-specific variability. These findings reveal a dissociation between task-level performance and metacognitive behavior in GPT-4, suggesting that its confidence reflects structural properties of its outputs rather than genuine internal uncertainty monitoring. Taken together, these findings suggest that GPT-4 lacks robust metacognitive abilities compared to humans, or at least that its metacognitive processes differ significantly from those of humans.