The Poet as Biological Language Model: On the Structural Resemblance Between Poetic Composition and Large Language Model Inference

This paper proposes a novel structural framework for understanding the interaction between human cognition and Large Language Models (LLMs) by drawing a rigorous analogy to formal poetic composition. Moving beyond metaphors such as “stochastic parrots” or “compression algorithms,” it identifies four shared dimensions between the skilled poet and autoregressive inference: (1) corpus formation, comparing experiential training data to textual datasets; (2) constrained generation, mapping prosodic requirements such as meter and rhyme onto sampling parameters such as temperature and top-p; (3) emergent semantics, analyzing how meaning arises bottom-up from formal constraints rather than top-down intent; and (4) the verifier function, equating the “poetic ear” with alignment and oversight mechanisms. The paper does not claim common mechanism, equivalence of cognition, or identity of intention between human and machine; rather, it advances a structural analogy intended to clarify how constrained generative systems produce, test, and refine meaning. It argues that “hallucinations” are not unique to artificial neural networks but are a general property of constrained generative systems—what poets recognize as the “false note.” By synthesizing 20th-century Russian Formalist theory with 21st-century transformer architecture, the paper proposes a “Poetic Posture” for AI governance. It concludes that effective AI alignment is not purely an engineering problem but an epistemological one, requiring a shift toward an apprenticeship model of calibrated trust in which the human operator reclaims the role of the unified generator and verifier. Keywords: Large Language Models (LLMs), AI Alignment, Human-AI Collaboration, Generative AI, Poetics, Epistemology, AI Governance, Emerging Technologies