The Misnomer of 'Natural Language Processing': A Critical Examination of Computational Linguistics' Foundational Category Errors

The field of artificial intelligence operates under a foundational category error encoded in the term Natural Language Processing. We argue that current large language models (LLMs) engage in neither natural processes, nor language in any philosophically or linguistically coherent sense, nor processing as understood in cognitive science. Instead, they perform high-dimensional statistical pattern matching on textual corpora, an extraordinary engineering achievement that remains persistently mislabeled. . Our argument unfolds in three phases: First, we demonstrate the empirical and architectural differences between human learning and transformer-based models. . Second, we examine the conceptual confusion that arises from treating understanding as substrate-independent (like flight) rather than substrate-dependent (like digestion). . Third, we investigate why language models work at all if they do not understand language. Drawing on Saussure's structuralism, distributional semantics, and the historical failure of symbolic AI, we analyse the structural conditions that make language amenable to statistical prediction and why this generates the appearance of understanding. . Through philosophical analysis (Searle, Dreyfus, structural linguistics), cognitive science, and computational history, we distinguish legitimate engineering accomplishment from ontological overreach. . We propose replacing Natural Language Processing with Statistical Text Synthesis to restore conceptual clarity and foster more responsible development. This is not a semantic quibble but a necessary intervention: what we call these systems shapes what we build, how we regulate them, and what we come to believe they can do.