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

This paper argues that the term "Natural Language Processing" encodes three category mistakes (Ryle 1949, pp. 16-18) that misrepresent what large language models actually accomplish, as evidenced by recent deployment incidents (Mata v. Avianca Airlines, Inc. 2023). Through philosophical analysis and technical examination, we demonstrate that these systems engage in neither natural processes, nor language in any philosophically defensible sense, nor processing as cognitive understanding. Drawing on evidence from phenomenology (In-der-Welt-sein in Heidegger 1962, p. 78; the corps propre, the lived body as the locus through which a world is given, in Merleau-Ponty 2012, p. 189), cognitive science (Dreyfus 1972, pp. 156-160; Searle 1980, pp. 417-424), and empirical studies of model behaviour (Brown et al. 2020, p. 8; Wei et al. 2022, pp. 3-7), we show that current transformer-based architectures perform statistical pattern matching on textual corpora through high-dimensional vector operations, a remarkable engineering achievement that has been systematically mislabelled. The evidence includes a learning efficiency gap between humans and models that runs in the opposite direction to the substrate advantage (Hart & Risley 1995, p. 132; Brown et al. 2020, p. 8), architectural conditions incompatible with what the term understanding names when applied to a Dasein in a world (Vaswani et al. 2017, pp. 5-6), and deployment incidents that reveal the consequences of this mislabelling (Mata v. Avianca Airlines, Inc. 2023, pp. 2-4). We identify three category mistakes: (1) mistaking artificial pattern-matching for natural cognition, (2) conflating text generation with linguistic practice, and (3) equating statistical processing with understanding. These are not semantic quibbles but category mistakes with implications for regulation, deployment, and human self-understanding (Hempel 1965, pp. 101-122). The paper concludes that accurate nomenclature, replacing "Natural Language Processing" with "Statistical Text Synthesis", carries weight in two directions at once: it gives the engineering work an honest description of its object, and it preserves the distinction between linguistic practice and the residue of linguistic practice that the misnomer collapses.