World Wide Models: Literary Tools for Cultural AI

The paper doesn't just argue that LLMs need cultural awareness (that's established). It specifically claims that comparative literature methods and translation theory are missing technical infrastructure, not just missing context. That's a claim about what kind of expertise should shape model architecture, not just training data.

This connects directly to the MSQA benchmark from last month, which showed that language fluency doesn't guarantee cultural competence and that inference-time fixes alone won't close the gap. The current paper goes further by naming what should replace those failed approaches: it's arguing that narratology and translation theory are the missing design layer. The MetaHOPE framework on metaphor translation also sits here, since metaphor is precisely the kind of culturally dense phenomenon that humanities scholarship handles but current benchmarks treat as a technical problem. Together these stories suggest the field is moving from 'we need to measure cultural gaps' to 'we need to rebuild how models are constructed to avoid those gaps in the first place.'

If major model developers (Anthropic, OpenAI, Meta) hire literary scholars or translation theorists into core model development roles within the next 18 months, that signals this framing is moving from academic proposal to industry practice. If they don't, the paper remains a humanities critique rather than a technical reorientation.