Whose Arabic? Diglossia, Dialect Gaps and Representational Justice in Arabic Large Language Models

Abstract

Arabic is spoken by hundreds of millions of people, but it is not one variety. Arabic speakers write and speak formally in Modern Standard Arabic and live much of their lives in regional dialects that differ widely from it and from each other. A new generation of Arabic-centric large language models, many built with public support in the Gulf states, promises to serve Arabic speakers in their own language. This article asks a question those projects have not fully faced: whose Arabic will these systems speak? Drawing on recent benchmark studies, it shows that Arabic models remain markedly weaker in dialects than in the standard language, that leading evaluation resources are built mainly around the standard language, and that some widely used dialect resources contain no data from the United Arab Emirates, Kuwait or Bahrain at all. The article then asks what kind of wrong such gaps involve. It distinguishes a distributive wrong (unequal quality of service), a recognitional wrong (failure to understand speakers on their own terms) and a standardising wrong (technology that quietly turns one register into the norm). Against both an "MSA-only" view and a demand for full dialect parity, it defends a principle of repertoire fidelity: systems should serve the full range of registers that their users actually move between, and should let users, not models, choose the register of a reply. It ends with proposals for evaluation, data partnerships, reporting and procurement, aimed at Gulf governments that fund national models.