Algorithms, language, and poetry: a phenomenological perspective

Summary

Algorithmic formalizations of language, such as those used in large language models, are not neutral tools but historically specific crystallizations of a more primordial field of embodied expression. Drawing on Merleau-Ponty's account of embodied speech and Heidegger's concept of technological enframing (Gestell), the authors argue that reducing language to optimizable signals risks suppressing its living, self-renewing capacity to generate new sense. The paper sketches phenomenologically informed criteria for language technologies that respect expressive openness, relational depth, and the historicity of signifiers, offering orientation for AI ethics debates.

Study at a glance

Abstract

This paper examines the algorithmic formalization of language through a phenomenological lens, engaging Martin Heidegger and Maurice Merleau-Ponty in dialogue with contemporary large language models (LLMs) and related AI systems. Instead of treating computationally modeled language as a neutral medium for information transfer, we argue that both formal logic and data-driven models are historically specific crystallizations of a more primordial field of embodied expression. The idea of the ”unity of language” refers to the dynamic, historically situated field of expressive possibilities within which multiple linguistic systems — natural languages, formal calculi, code, poetic language — emerge, sediment, and transform. Drawing on Merleau-Ponty’s account of embodied speech, we reconstruct language as a living, self-renewing medium whose unity lies in its ongoing capacity to generate new sense. Heidegger’s analysis of technological ”enframing” (Gestell) and his reflections on ”traditional language” then allow us to interpret algorithmic conceptions of language as powerful, but critically informing of the existential risk of reducing speech to optimizable signals within the wider field of linguistic life. We confront these insights with current developments in AI, including LLMs, embodied AI, and enactive or 4E approaches to cognition. We conclude by sketching phenomenologically informed criteria for language technologies that respect expressive openness, relational depth, and the historicity of signifiers, and indicate how such criteria can orient debates in AI ethics.