Cultural Collapse: Toward a generative formalism for AI cultural production

This paper examines systematic patterns of idealization in large language model outputs through computational analysis of over 15,000 AI-generated poems and artificial bibliographic data. The study reveals and theorizes ‘cultural collapse’—the tendency of LLMs to generate cultural content that is more formulaic and idealized than can be observed in any historical period. Analysis of rhyme patterns shows that models produce formally conservative verse at rates that exceed even the most traditional historical periods. This bias persists even when models are explicitly instructed against traditional forms and cannot be explained by training data composition, suggesting deep computational tendencies toward idealization. Extending beyond poetics, parallel patterns emerge in historical domains: when prompted to generate historical publication data, models systematically produce demographic distributions that obscure well-known exclusion patterns, creating revisionist narratives where marginalized authors were published at rates far exceeding historical reality. The study identifies instruction tuning as one contributing mechanism, with models fine-tuned to be helpful assistants showing significantly greater ‘idealization’ than base models. These findings suggest that cultural collapse operates through a computational logic that privileges satisfaction over frustration, regularity over variation, and conformity over contradiction. As generative systems become ubiquitous in cultural production, their idealizing tendencies threaten to flatten cultural diversity and historical complexity, requiring new critical frameworks for understanding computational mediation of cultural transmission.