Language Models and the Flattening of Sociolinguistic Variation

This paper asks what happens when large language models (LLMs) meet non-standard dialects. We focused on two simple questions. First, how do these models react when a prompt is written in a regional or social variety instead of the standard? Second, do they actually keep any of those features, or do they smooth them out into the standard form? To test this, we put together a small dataset. On the English side, we used African American English, Appalachian English, and some regional UK forms. On the German side, we included Bavarian and Swiss German. Each prompt was entered into a state-of-the-art model several times. We kept all of the replies. Our method was fairly straightforward. We coded the outputs by hand: did the dialect form stay, did it get changed, or was it treated like an error? Then we counted. How many features were in the input? How many made it to the output? It is a simple approach, but it let us see both the fine details and the larger picture. The study draws on variationist sociolinguistics and the sociology of language. Labov and Eckert show us that variation is structured and meaningful. Bourdieu’s idea of linguistic capital explains why certain varieties carry more weight. And of course, the ideology of the standard language is always in the background—the belief that there should be one correct form, and everything else is less legitimate. The results were clear, and honestly a little disappointing. The model almost always shifted dialect into the standard. African American English was answered in Standard English. Swiss German prompts came back in Standard High German. Sometimes the model even hinted that the input was “wrong.” In other words, it leveled out the variation. What stood out to us was not only the frequency of this pattern but the way it repeats social hierarchies. If the model keeps erasing dialect features, it reduces the visible space of linguistic diversity online. Language is not only about getting a message across—it is also about identity, community, and belonging. When models flatten dialects, they also echo old power relations: the standard survives, the others are pushed aside. This matters for the future of AI. If LLMs continue down this path, they risk reinforcing inequality instead of opening space for more voices. We argue that models should be trained and evaluated with sociolinguistic variation in mind. Otherwise, the cost of convenience will be the loss of diversity.