What do LLMs remember? A linguistic-sociological framework for studying and using language technology

Language technologies, especially those related to AI and machine learning, have grown increasingly popular as objects and methods of study for sociologists. We contend that to understand these objects and their place in our social world, we must attend to a linguistic sociology. In the context of large language models, for example, language -text, specifically- acts as both the input and output interface between person and machine. Yet text is itself a language technology, and carries with it affordances and constraints that remember long histories of social relations. Quilting together literature from sociology, computational linguistics, science and technology studies, and linguistic anthropology, we ask: what do language technologies remember, and how can sociologists take these memories into account? To answer this question, we offer a three-part analytical framework to examine how the memories of language technologies are both constructed by and constructing our social world. Drawing on two examples of use of language technologies in college admissions and in scientific peer review, we apply our framework to demonstrate how language technologies can mediate social actions, processes, and relations.