Against the Binary Question: An Anti-Semantic Decomposition of LLM Understanding

Whether large language models understand is often treated as a binary question. This paper argues that the question is analytically underdetermined because it compresses non-identical positions into the single predicate “understands.” The analysis distinguishes eight core positions: input configuration, processing configuration, output configuration, local followability, semantic judgment, explanatory evaluation, attribution of understanding, and institutional acceptance. These positions are not claimed to form an exhaustive taxonomy or a necessary temporal sequence. They constitute a non-redundant core map for the present debate: each has a distinct analytical target, can vary without requiring corresponding variation in all neighboring positions, and generates different inferential consequences when identified with another position. The framework further distinguishes kinds of understanding from positions at which evidence for, evaluations of, and attributions of such kinds occur. This separation shows why pluralist accounts of understanding and mechanistic indicators do not eliminate the need to distinguish evidence, criteria, attribution, possible possession, and institutional stabilization. Mechanistic evidence may materially strengthen an understanding-attribution, but a mechanism, evidence for that mechanism, the criterion connecting it to understanding, the scope of the resulting attribution, and the attributed property are not thereby identical. The paper develops a prospective test for analytical compression by requiring any transition between positions to identify its evidential basis, licensing criterion, target claim, and scope, and demonstrates the test on a worked case rather than presenting it only in schematic form. It also extends the framework to multi-turn interaction, retrieval-augmented generation, and tool-using systems, where success may belong to a configured system rather than to a base model considered in isolation. Fluent or explanatorily useful output does not by itself establish understanding, while the absence of self-proving evidence does not make attribution arbitrary. Anti-semantic decomposition therefore precedes rather than replaces substantive affirmative or negative theories of LLM understanding.