Stewardship, Not Extraction: Large Language Models and Endangered Language Revitalization

Abstract

This article examines the rapidly expanding application of large language models, automatic speech recognition, text-to-speech synthesis, and machine translation to the documentation and revitalization of endangered and Indigenous languages. It argues that the encounter between data-hungry neural architectures and the world's smallest speech communities is structurally fraught: the very statistical logic that makes contemporary language technology powerful presupposes the abundant text that endangered languages, by definition, lack, while the dominant practice of harvesting linguistic material as freely available "data" reproduces the extractive dynamics that contributed to language endangerment in the first place. Drawing on computational scholarship on low-resource languages, on the Indigenous data sovereignty literature, and on concrete cases including te reo Maori, Welsh, and Cherokee, the discussion situates technical questions of scarcity and scale within ethical questions of agency, ownership, and benefit. It contends that frameworks such as the CARE Principles, OCAP, and community-authored licensing offer a constructive path beyond a sterile opposition between technological optimism and refusal. The article concludes that artificial intelligence can meaningfully support revitalization only when subordinated to community-led documentation and, above all, to the intergenerational transmission that remains the decisive measure of a language's future.