Latent linguistics: the conception of language in large language models

The emergence of large language models (LLMs), such as GPT, Mistral, and Llama, marks a significant moment in the history of language technologies. Initially designed as tools for predicting and generating text, these models have quickly become the foundation for a wide range of applications, including dialogue systems like ChatGPT, code generation tools, writing assistants, search engines, and educational platforms. Their capabilities have sparked intense debate—not only about their practical uses but also about the assumptions they embody regarding language, meaning, and cognition. In many respects, LLMs force a reengagement with long-standing philosophical questions about the nature of language and its relationship to thought and the world. At the core of LLMs lies a relatively simple objective: Given a sequence of words, predict the most likely continuation. This statistical task, scaled to unprecedented data volumes and computational resources, has produced systems capable of producing fluent, coherent, and contextually sensitive text. Yet, despite their apparent sophistication, LLMs operate entirely within the domain of language itself. They do not perceive, act, or refer to the external world. Rather, they model language by detecting patterns of cooccurrence in large textual corpora, generating responses by recombining elements of prior discourse. As such, they instantiate a particular conception of language—one in which meaning arises not from reference or intention but from internal relations among signs. This vision has important intellectual antecedents. Structuralist and poststructuralist thinkers such as Saussure, Barthes, and Foucault argue that language is not a transparent medium for representing reality but a structured system that shapes thought and subjectivity. They emphasize that meaning is not inherent in words but emerges from differential relations, discursive conventions, and cultural codes. While these were originally philosophical and critical interventions, they now resonate in unexpected ways with the functioning of LLMs—systems that produce language without any stable referent, authorial agency, or grounding in lived experience. This structuralist affinity, however, is only a starting point. The goal of this chapter is not to situate LLMs within a particular theoretical tradition but to examine the model of language they enact. What kind of representation is involved in a system that learns exclusively from textual data? What semantic capacities, if any, can be attributed to a model that lacks perception, embodiment, or communicative intent? And what do these systems reveal about the limits of language-based modeling more generally? To explore these questions, we begin with a discussion of structuralist and distributional approaches to meaning, before turning to more recent debates about grounding, reference, and interpretability. The aim is to understand how LLMs manage coherence without understanding, and what their successes and failures can teach us about the nature of linguistic knowledge, the epistemology of statistical modeling, and the boundaries between linguistic form and semantic content.