Begus applies world literature tools to monolingual LLMs

A short humanities paper posted to arXiv this week is worth flagging not because it announces a new model, but because it names something practitioners increasingly bump into and then set aside. Nina Begus's essay, "World Wide Models: Literary Tools for Cultural AI", argues that LLMs stage "a new form of cultural encounter that is massive, automated, and monolingual."

The move is straightforward. The tools literary scholarship has developed for handling cross-cultural texts, including comparative reading, narratological and poetic analysis, critical theory, world literature, and translation, are, in her framing, the same tools we now need for building AI that is not tacitly stuck inside one linguistic worldview. She proposes a layered framework aimed at more nuanced textual models and pluralistic interpretations, and she pulls specifically on world literature approaches, naming macrostructure, circulation, and untranslatability as ways to think about what she calls global AI textuality.

Why this matters if you are not a literary scholar: the debate over multilingual capability has mostly been framed as a data-and-benchmark problem. Begus is arguing that framing is thin. Structural monolingualism, in her account, is not just a coverage gap that shrinks as you add tokens in other languages, it is a stance embedded in how a system reads and generates meaning, and critical theory has spent a long time developing vocabulary for exactly that.

The honest caveat is that this is a 15-page essay, forthcoming in MFS Modern Fiction Studies in 2027, not an engineering paper. What the abstract does not give you is a worked example, a benchmark, or a critique of any specific model, so whether any of this vocabulary lands inside actual model development is an open question.

The direction is the part worth watching. If evaluation practice for multilingual and cross-cultural AI starts borrowing from world literature rather than only from NLP, the definition of a good model along that axis gets more interesting for the labs, and more useful for everyone reading their outputs in a language other than English.