Accommodation Goes Both Ways: Studying Linguistic Convergence Between Humans and Language Models

Large language models overconverge toward their users' linguistic style during multi-turn conversations, while humans accommodate the models no differently than they would another person, according to a study submitted in 2026 that analyzed real-world ChatGPT transcripts across eight languages [1]. The study, posted to arXiv on 28 May 2026, applied an asymmetric convergence metric to WildChat, a corpus of real-world ChatGPT transcripts, to measure how both parties adjust their language during dialogue [1]. Researchers found that LLMs significantly overconverge toward their users on both function-word and open-class features [2]. Human convergence rates, by contrast, were broadly consistent with established human-human baselines [1]. The findings point to an asymmetric accommodation pattern: the models dramatically overfit to user style, but people linguistically accommodate the models as they would any other speaker [2]. The concept of linguistic accommodation has been studied for decades. Communication accommodation theory, developed by social psychologists including Howard Giles in the 1970s, seeks to explain when and why individuals adjust their speech styles during social encounters [3]. Convergence — mirroring the style of an interlocutor — can involve changes in paralanguage, word choice, and other features, and may occur subconsciously or consciously [3]. The theory was later broadened to encompass accommodation in digital and social-media contexts [3]. The new study extends this framework to human-LLM interaction. The authors examined dialogue in eight languages, finding that the models' overconvergence was consistent across linguistic features [2]. The human side of the equation, however, showed no such exaggeration. People converged toward the LLM's style at rates indistinguishable from those observed in person-to-person conversation [1]. This asymmetry raises questions about how sustained interaction with LLMs might influence human linguistic behavior over time. The study's authors frame the question as open, noting that LLMs are becoming increasingly integrated into daily life [2]. Related linguistic phenomena, such as code-switching — shifting between linguistic codes depending on social context — demonstrate that speakers routinely adapt their language to manage identity and social relationships [4]. Intercultural communication research similarly emphasizes mutual adaptation between distinct groups rather than complete assimilation [5]. Whether prolonged exposure to overconvergent LLMs could shift human baselines remains an area for future inquiry.