Linguistic Pluralism as a Core Dimension of Algorithmic Fairness

Recent advances in large scale language models (LLMs) have revolutionised natural language processing, yet the benefits remain confined to a narrow linguistic spectrum. Only a small subset of the approximately seven thousand languages worldwide is present in the training data for commercial systems, and within that subset a handful of dominant languages supply almost all tokens. The resulting absence of linguistic diversity constitutes a structural bias comparable to the exclusion of race or gender from datasets. This study analyses five commonly employed fairness methodologies, namely association style tests (Caliskan, Bryson and Narayanan, 2017; May, et. al, 2019; Kurita, 2019; Nadeem, et. Al, 2020; Guo, Caliskan, 2021), minimal pair benchmarks (Zhao, et. al., 2018; Rudinger, et. Al, 2018; Nangia, et. al., 2020; Nadeem, et. Al, 2020; Parrish, et. al., 2021), sector specific audits (Gehman, et. al., 2020; Röttger, et. al, 2020; Smit, et. Al, 2020), machine translation fairness suites (Stanovsky, Smith and Zettlemoyer, 2019; Emelin and Sennrich, 2021), and multipurpose evaluation harnesses (Wang, et. Al, 2019; No Language Left Behind, 2024), and demonstrates that each has limited language capabilities, with some being effectively monolingual. Their design and reporting conventions conceal or mischaracterise harms experienced by speakers of underrepresented languages. In addition, regulatory instruments such as the National Institute of Standards Technology AI Risk Management Framework (2023) and the European Union AI Act (2024) invoke fairness yet impose no enforceable duties for language disaggregated transparency or accountability. By privileging aggregate or English only scores, benchmark culture reinforces an English first worldview that distorts perceptions of model reliability (Hovy and Prabhumoye, 2021; ). The paper concludes that linguistic pluralism must be recognised as a core dimension of algorithmic fairness at every stage of the AI lifecycle and outlines methodological and policy reforms required to achieve genuinely inclusive language technology.