AI text transformation and linguistic inequality: how computational mechanisms mediate multilingual academic knowledge production

Abstract AI-based text transformation systems have become mediators of multilingual academic knowledge production. This study argues that the computational mechanisms underlying these systems, including training data distribution, curation, tokenization, and optimization, do not extend multilingual capability neutrally: they structurally privilege certain languages, genres, and epistemic conventions, reproducing the linguistic hierarchies that applied linguistics has long documented in institutional practice. Through a conceptual-interpretive synthesis of 24 studies, the study develops a three-layer framework of transformation processes, epistemic transfer, and institutional mediation. The central argument is that inequality in AI-mediated multilingual contexts may operate not only through representational gaps but through transformation itself, and the framework identifies three dynamics through which this is proposed to occur: the selective preservation of high-frequency content, the contextual loss of embedded meaning, and the normative realignment of outputs towards dominant epistemic conventions. AI systems are unlikely to dissolve language hierarchies; the analysis suggests they tend to embed them computationally. The study reconceptualises multilingualism in AI not as a property of model coverage but as an outcome of transformation.