Toward a Culturally Embedded Arabic LLM: Localizing Linguistic AI for the Arab World

Abstract

Arabic large language models (LLMs) increasingly support communication, education, public services, and knowledge access, yet fluent Arabic generation does not necessarily constitute culturally appropriate understanding. This paper examines the problem of cultural and pragmatic misalignment in contemporary Arabic-capable models and proposes a framework for culturally embedded Arabic AI. Cultural embeddedness is operationalized through four interrelated capacities: Arabic-native semantic representation, register awareness, cultural inference, and dialectal continuity. Using a mixed-methods design, the study introduces the Culturally Sensitive Test Suite for Arabic (CSTS-Arabic) and a four-dimensional Cultural Quotient (CQ-Score) for evaluating contextual appropriateness, dialectal consistency, cultural-reference integrity, and pragmatic success. The study compares multilingual models, Arabic-focused models, and a proof-of-concept Culturally Aware Language Model (CALM). The reported findings indicate a persistent cultural competence gap in multilingual systems and improved performance when dialect-aware representations, structured cultural retrieval, and pragmatic attention are integrated into the modeling pipeline. The paper further develops an Arabic-first architectural proposal, a tiered corpus strategy, and ethical safeguards concerning representation, doctrinal plurality, epistemic authority, evaluation bias, and resource justice. Rather than treating culture as an optional layer of factual knowledge, the study argues that it should be modeled as a constitute part of linguistic meaning. The proposed framework therefore positions culturally embedded Arabic LLMs as both a technical and governance challenge and offers a pathway toward more reliable, pluralistic, and regionally accountable Arabic AI.