World Wide Models: Literary Tools for Cultural AI

The Literary Lens on AI’s Cultural Monolingualism

A new arXiv paper positions large language models as engines of a “massive, automated, and monolingual” cultural encounter, drawing on literary disciplines—comparative reading, narratology, and critical theory—to diagnose this phenomenon. The authors argue that literary tools, historically used to analyze how texts negotiate cultural struggles, offer a critical framework for understanding LLMs’ homogenizing effects on global culture.

The core observation is sharp: LLMs train on predominantly English-language internet text, then serve as universal interfaces for users worldwide. This creates a feedback loop where non-English cultural expressions are either absent, translated into flattened approximations, or actively suppressed by safety filters and tokenization biases. Literary analysis, with its long history of examining how narratives encode power, perspective, and cultural specificity, provides a vocabulary to name what is being lost—not just linguistic diversity, but the narrative structures, poetic rhythms, and contextual meanings that constitute distinct cultural identities.

Why this matters beyond academia. The paper implicitly challenges the prevailing AI industry assumption that scaling data and compute will solve cultural representation. It suggests that the problem is not merely one of missing data (e.g., adding more non-English books) but of structural monolingualism embedded in model architecture, training objectives, and evaluation metrics. For AI practitioners, this reframes “alignment” and “safety” as cultural curation problems, not just technical ones. A model that cannot tell a story in the narrative tradition of a specific oral culture, or that flattens a poetic form into prose, is not just inaccurate—it is performing a kind of cultural erasure. Implications for AI developers and deployers. First, multilingual training data is necessary but insufficient. Practitioners must audit models for narrative diversity, not just lexical coverage. Second, literary theory offers concrete methods—close reading, comparative analysis, genre studies—that can be operationalized into evaluation benchmarks. Third, the paper suggests that current RLHF and preference tuning may inadvertently reinforce cultural homogeneity by optimizing for a “universal” user preference that is, in practice, Western and English-dominant. Finally, for product teams building global-facing AI, the takeaway is clear: cultural sensitivity requires deep engagement with local narrative traditions, not just translation APIs.

Key Takeaways

• LLMs are creating a massive, automated cultural encounter that is structurally monolingual, flattening non-English narrative traditions and poetic forms.
• Literary disciplines provide a rigorous analytical toolkit—comparative reading, narratology, critical theory—for diagnosing and mitigating this cultural homogenization.
• AI practitioners must move beyond token-level multilingualism to evaluate models on narrative diversity, genre fidelity, and cultural context preservation.
• Current alignment and safety techniques may inadvertently reinforce English-centric cultural norms, requiring new benchmarks grounded in literary analysis.