This paper introduces Poetry for AI, a literary genre and performance practice in which the intended reader of the text is a large language model rather than a human being. The genre is distinguished not by its subject matter, which need not concern artificial intelligence at all, but by its addressee and by a compositional strategy aimed at the model's manner of reading: the sequential, expectation-driven construction of meaning from an accumulating context. Where conventional poetry offers a message to be decoded, these poems are constructed to act on the act of reading itself. The paper defines the genre through a three-axis taxonomy, describes its compositional principles in the vocabulary of cognitive poetics and reader-response theory, specifies a two-message elicitation protocol designed as a control against confabulation, and frames the documented model responses as a form of performance and relational art. It situates the genre among concrete poetry, the OULIPO, codework, conceptual writing, and ergodic literature, and argues that its novelty lies in the displacement of the addressee from the human to the machine. The paper makes no claim regarding machine sentience or feeling, and it claims no access to model internals; its interest is the recurring and describable character of the responses the genre elicits, and the new addressee-relation it opens. A founding corpus of three poems accompanies the paper.
Dmitriy Pos (2026) studied this question.