Information Dynamics: Toward a Unified Interpretation of Stability in Biological and Artificial Cognitive Systems

This paper proposes a unifying framework for understanding intelligence as a process governed by the dynamics of information under constraint. Rather than focusing on performance or generative capability, the analysis begins with biological cognition, where stability emerges despite limited memory capacity, reconstructive recall, and incomplete information. We argue that both biological cognition and artificial systems can be understood as systems subject to information dynamics, where stability emerges through regulation under constraint. Drawing on findings from neuroscience and cognitive science - including memory reconstruction [1][2], consolidation during sleep [3][4], executive control [7][8], and socio-contextual framing [6] - we identify key mechanisms that regulate information dynamics. These mechanisms are interpreted as stabilizing processes that maintain coherence over time. Artificial systems, particularly large language models (LLMs), are then analyzed through the same lens. Their characteristic instabilities - such as hallucinations, drift, and inconsistency - are argued to arise not from insufficient capability, but from the absence of intrinsic regulatory structures. For the sake of clarity, we acknowledge that some limited stabilization mechanisms emerge from architectural constraints of artificial systems - preset declarative and procedural memory acquired at training. We introduce a minimal model of information dynamics to formalize these observations and argue that external control frameworks can function as prosthetic regulatory layers. These compensate for missing stabilization mechanisms, suggesting that intelligence is best understood as a property of regulated systems rather than autonomous agents.