Large language models and creativity

The rapid development of large language models (LLMs) has revived longstanding debates about the nature of creativity and the possibility of machine-generated art. This paper examines the case of poetic generation as a particularly revealing testbed for evaluating artificial creativity. Drawing on historical experiments in computational poetry and contemporary philosophical theories, it analyzes how current models produce verse and assesses the extent to which such outputs can be considered genuinely creative. While LLMs demonstrate a high level of formal competence—successfully reproducing poetic structures, stylistic conventions, and thematic coherence—their outputs remain grounded in large-scale statistical recombination of existing textual patterns. As a result, they primarily exhibit combinational and exploratory forms of creativity, while lacking the capacity for transformational creativity that would involve redefining the underlying conceptual space. Empirical studies further show that although AI-generated poetry can be perceived as comparable to human writing in blind evaluations, it tends to display stylistic homogeneity, limited conceptual depth, and a reliance on conventional tropes. Beyond technical performance, the paper highlights broader challenges related to aesthetic evaluation, cultural bias in training data, and the attribution of authorship and creative value. It argues that, rather than constituting evidence of autonomous machine creativity, LLM-generated poetry is best understood as the product of sophisticated pattern generation, whose significance lies as much in its role within human–machine interaction as in the artifacts it produces.