Gendered asymmetries in AI stories in English and Spanish

Abstract

Creative texts produced wholly or partly by large-scale generative language models are increasingly circulating within literary markets and cultural institutions. Research on AI-generated texts has largely examined bias at the level of words and images; however, small-scale statistical co-occurrences in AI-output can generate macro-scale narrative phenomena, where narrative functions intersect with stylistic texture, for example through voice, focalisation, reliability, or omission. This article argues that literary criticism, and New Formalism in particular, offers a necessary methodological framework for identifying and historicising these stylistic asymmetries. A background premise, working against the current techno-utopian ethos that improving algorithms requires more algorithms, is that close, slow reading is necessary to addressing the ethical complexities in literary form, the unsettled nature of language, and the context-specific nature of bias, harm, and oppression (Jackson and Courneya 2023, p. 62). Through a new formalist reading of English- and Spanish-language short stories generated by ChatGPT-5, I demonstrate that distributional regularities in language modelling scale upward, producing gendered and culturally normative stylistic patterns. By shifting the analysis of AI creativity from representation to form, I hope to indicate how literary methods can illuminate narrative-specific dimensions of bias in generative AI output.