Low-resource language

Traditional definitions of language resourcedness fail to capture the structural inequalities embedded in LLMs’ pre-training data. While historical classifications rely heavily on sociopolitical vitality, artifacts, and human resources, a language’s real-world robustness no longer guarantees equitable algorithmic representation. To address this disparity, this paper proposes a novel tripartite framework that combines four established exogenous criteria with a relative endogenous quantitative metric: a language must hold a dominant or co-dominant statistical share of an LLM’s pre-training dataset to be classified as high-resource. Under this new paradigm, languages with strong societal infrastructure but marginal AI data representation, such as Italian and German in LLMs like GPT-3, are accurately reclassified as few-resource, while those lacking both real-world and digital infrastructure remain low-resource. Despite industry trends toward data opacity, adopting this precise terminology equips researchers with the tools to expose engineered linguistic disparities, challenge the false equivalence of traditional labels, and advocate for genuinely equitable, multilingual AI systems.