What certainty and other metacognitive representations in cognitive development reveal about the future of AI models

Large language models (LLMs) are increasingly used for learning, information seeking, decision-making, and interaction, making their ability to represent the limits of their own knowledge increasingly important. Existing research on LLM metacognition has largely evaluated whether model confidence tracks objective performance. We argue that confidence-accuracy calibration alone provides an incomplete characterization of metacognition. Drawing on research from developmental psychology, we show that certainty emerges early, becomes increasingly calibrated through experience, can incorporate information beyond current performance, and can guide subsequent cognition and learning. We use these findings to propose three directions for studying metacognition in artificial systems: examining how metacognitive abilities change under different training conditions, characterizing what information model certainty represents, and testing whether certainty regulates subsequent learning and behavior. This developmental perspective motivates studying LLM metacognition as a representational and functional capacity.