Language as a Low-Bandwidth Interface: Capacity Constraints and the Reconstruction of Meaning

This paper presents a framework that treats natural language not as a mirror of the world or as a container of meaning, but as a low-bandwidth interface through which finite agents reconstruct high-dimensional cognitive states. The starting point is the capacity gap among the world W, cognition C, and language L, together with the double irreversibility involved in W → C and C → L. This gap is not a reduction of meaning to bits; rather, it is an order-of-magnitude comparison of bandwidth and degrees of freedom that constrains strong mapping theories and content-transfer theories. Language does not function as a medium that carries content intact, but as a low-bandwidth pointer that guides reconstruction on the recipient’s side. From this perspective, the paper treats linguistic expressions as public pointers p and distinguishes the intra-individual operable structure Dᵢ(p) from the public relational structure R. This distinction clarifies the difference between being able to manipulate definitions and explanations fluently and being able to identify, predict, and revise the target. NULL pointers, meaning inflation, decoder synchronization, explanatory clarity, social reality, and hallucination in LLMs can all be explained as problems of reconstruction and stopping conditions. Furthermore, by positioning institutional and fictional concepts as R-dominant pointers, the framework explains why LLMs can be strong in law, policy, and fictional worlds while still requiring external constraints in physical manipulation, measurement, experimentation, and formal verification. In this restricted, language-mediated sense, the framework recasts AI alignment not merely as the generation of correct answers, but as a design problem involving grounding, verification, stopping, and decoder synchronization. More specifically, it treats alignment as a problem of controlling how goals, values, evidence, and responsibility are reconstructed from natural-language pointers.