The Knowledge Density Hypothesis of Confabulation: A Unified Framework for Understanding Confabulation Across Biological and Artificial Intelligence

This paper proposes the Knowledge Density Hypothesis of Confabulation (KDH): the claim that confabulation confidence in any reconstructive memory system — biological or artificial — is primarily determined by knowledge density, a measure of how completely stored patterns cover the space of possible queries. When density is low, retrieval visibly fails and uncertainty signals fire. When density is extreme — as in large language models trained on internet-scale data — virtually every query activates sufficient patterns to produce a plausible construction, eliminating the structural conditions that generate epistemic humility. The paper synthesizes evidence from cognitive psychology (constructive memory, metacognition, Dunning-Kruger), AI interpretability (hallucination-associated neurons), and clinical neuroscience (orbitofrontal reality filtering), arguing that the same functional dynamic underlies confabulation across substrates. The central falsifiable prediction — the Inverted U-Curve of Calibration — holds that calibration improves with knowledge up to moderate expertise, then degrades at extreme density.