A Sociolinguist's Look at the “Language” in Large Language Models

Large language models (LLMs) work by training text-generating systems through the use of digitized corpora that make natural language available in datafied form for natural language processing. LLMs are rooted in these culture-specific, socio-historically conditioned understandings of language, inaugurated during the imperial age of national print cultures and further consolidated by computing culture itself. However, understanding languages as digital data sets whose patterns can be probabilistically reproduced is a simplistic reduction of an already reified linguistic epistemology. This article discusses the sociolinguistic implications of LLM design and implementation.