Addressing Historical and Contemporary Inequalities in Low-Resource Languages through the Lens of Serbian

Abstract

The development of large language models prioritizes high-resource languages such as English, while most of the world's 7,000 languages remain underrepresented in AI. Multilingual pre-training and cross-lingual transfer have produced uneven results, favoring languages structurally closer to English and transferring linguistic and cultural biases from the training data into low-resource language (LRL) outputs, creating a double disadvantage in which LRLs are simultaneously underrepresented and misrepresented. This paper argues that developing language technologies for LRLs requires not only technical solutions, but also contextualized ethnographic inquiry into the historical, sociocultural, and institutional conditions that shape their digital future. Using Serbian as a case study, it draws on semi-structured interviews with ten scholars and NLP practitioners working across disciplines. Findings identify barriers common to many LRLs, including limited corpora, insufficient digitization, restrictive copyright regimes, and partial institutional support, which contribute to reliance on low-quality web data and multilingual models that reproduce bias. The paper outlines an alternative approach centered on native language models, interdisciplinary collaboration, inclusive datasets, and transparent data practices, with sustained involvement of humanities and social science scholars across data selection, evaluation, and governance, so that sociocultural and ethical concerns are built into language technologies from the outset rather than treated as post hoc technical interventions or matters of regulatory compliance.