Structures of Signs without Cultural Knowledge. Narrative Competence and Cultural Specificity in AI-Generated Fiction

Abstract

This article argues that narrative competence and cultural competence are independent variables in AI-generated fiction: a text can be mimetically vivid, synthetically controlled, and thematically coherent while getting its cultural and historical substance precisely wrong. To demonstrate this, it compares two systems – ChatGPT, a globally trained commercial model, and PLLUM, a Polish academic model – across a corpus of ten prompts requiring short stories set in post-1939 Polish contexts, with the culturally decisive reference embedded in the situation rather than stated. The method combines two instruments applied in sequence: James Phelan’s tripartite (mimetic, synthetic, thematic) model, adapted posthumanistically to treat AI outputs as “structures of signs” rather than rhetorical acts; and a cultural competence rubric assessing factual accuracy, nuance, contested material, and depth of embedding. Two case studies – a 1967 Kraków scene and a contemporary scene turning on the rapper Mata’s Patointeligencja – yield a typology of four failure modes: Generic Approximation, Emotional Inversion, Content Substitution, and Perspectival Misreading. The discussion proposes a distinction between cultural familiarity and historical situatedness, and argues that the transparency asymmetry between academic and commercial AI is constitutive of the question of whose cultural knowledge these systems encode.