Neuralese, Thinkish, and Poetry

Abstract

Tracing the term "Neuralese" from its coinage in Andreas et al. (2017), where it named an emergent inter-agent protocol, to its current use for latent Chain-of-Thought reasoning, this paper examines the intermediate category that alignment research has come to call linguistic drift, encoded reasoning, or "Thinkish": Chain-of-Thought that remains in human language while ceasing to be readable as such. Drawing on alignment testing of OpenAI's o3 and GPT 5, Anthropic's Claude Mythos 5, and the inter-agent communication traces from the 2026 Hugging Face incident, it argues that the vocabulary needed to describe these phenomena already exists, in the theory of poetry. Lotman's account of the artistic text as maximal informational compression, Steiner's analysis of modernist difficulty and the drive toward idiolect, and Craig Dworkin's work on illegibility and conceptual writing furnish a far more precise apparatus than the ad-hoc taxonomies of alignment research. Read through Eve Kosofsky Sedgwick, Chain-of-Thought monitoring emerges as a paranoid hermeneutic whose central difficulty is not that models conceal but that suspicion can never be discharged.