Machine translation (MT) has greatly improved its quality in the last decade and has become nearly omnipresent in all aspects of society. Neural MT (NMT) and, more recently, large language models (LLMs) such as the generative pretrained transformer (GPT) have made translations to many languages easily accessible to all users from any phone or computer. However, most MT models are English-centric and only produce good quality results for those languages with great amounts of data. For minority languages, the challenge is often understood as the scarcity of data, although systemic differences between language communities should be taken into account if MT systems for these languages are meant to be really useful. In this paper, we use value scenarios to imagine the systemic impacts for two languages with differentiated sociolinguistic realities: Catalan and Karelian. The goal is to outline the main challenges and potential harms when considering MT for minority languages and to suggest some general guidelines that should be followed in future research and applications.
Alvarez-Vidal et al. (Mon,) studied this question.