Abstract
As large language model (LLM)-based chatbots become increasingly embedded in educational settings, there is a growing need to understand how learning unfolds interactionally through student–AI dialogue. This study examines revoicing, a discourse strategy theorized by O’Connor and Michaels as a means through which teachers take up and reformulate student contributions to support disciplinary alignment and development of academic voice. While revoicing has been extensively studied in classroom talk, it remains under-theorized in the emerging context of human–AI educational interaction. In this study, we reconceptualize revoicing in student–AI chat as a reciprocal process of linguistic “threading,” whereby ideas are taken up and rearticulated through identifiable lexical and semantic relations across turns. Drawing on a Bakhtinian dialogic lens, we analyzed chatlogs from 22 secondary school students engaging with a customized GPT-4-based chatbot designed for science inquiry. Using adjacency pairs as the unit of analysis, we conducted a two-phase lexical–semantic content analysis of revoicing relations, including lexical repetition and word-family variation, as well as synonymous, specific–general, part-whole and implied relations. Findings indicate both the chatbot and students took up and extended one another’s ideas through multiple ways of revoicing. These interactional patterns extend existing conceptualizations of revoicing and demonstrate how AI systems can function as customizable revoicing agents in educational dialogue. The study calls for further AIED research on how AI-mediated revoicing shapes learners’ participation, agency, and socialization into academic roles, practices and identity.
