On the Inability of Language Models to Stub Their Toe: Referential Collapse, Semantic Inflation, and the Disconnect Between Linguistic Competence and Grounded Understanding

This paper examines a structural mismatch between the linguistic conventions rewarded in academic writing and the referential function language originally evolved to serve. Using a single physical object as a grounding example throughout, the analysis demonstrates how verbosity and elaboration are systematically rewarded while referential clarity is penalized in the corpora used to train large language models. Tracing the philosophical and empirical foundations of this mismatch through the symbol grounding problem, embodied cognition, and developmental research on joint attention, the paper argues that language models inherit an inverted training signal that optimizes for description rather than reference. As a result, such systems can generate increasingly elaborate linguistic output while remaining disconnected from the physical world to which language ultimately refers. The analysis concludes that understanding occurs exclusively in the human reader rather than in the system producing text, with implications for how language models are deployed in educational, interpretive, and decision-making contexts.