Language is primarily a tool for communication rather than thought (2024) [pdf] (gwern.net)

A new Perspective (Fedorenko, Piantadosi & Gibson, Nature 2024) argues that human language evolved and functions primarily as a tool for communication, not as the machinery of thought. The authors synthesize two decades of neuroscience, cross-linguistic corpus work and information‑theoretic analyses to make three core points: (1) a dedicated left‑hemisphere “language network” encodes abstract, modality‑independent linguistic representations and is exquisitely sensitive to word meanings and syntactic dependencies; (2) there is a double dissociation between language and many forms of thought—people with severe aphasia or when performing mathematical and logical reasoning often show intact non‑linguistic cognition despite disrupted language processing; and (3) structural features of languages and empirical regularities are better explained by pressures for efficient information transfer than by a requirement that language be the substrate of thought. For the AI/ML and cognitive‑science communities this reframes how we interpret linguistic representations in models and brains: language is optimized for transmission and cultural transmission, not necessarily for implementing general reasoning. Practically, the paper supports modular architectures where symbolic or distributed non‑linguistic representations can support reasoning independently of surface language systems, cautions against equating internal token sequences with “thought,” and encourages using information‑theoretic and neural‑mapping tools to distinguish communicative encoding from core cognitive computations.