Conversation by design: repair and interactional competence in L2 talk with an AI speaking application

Abstract AI-powered speaking applications invite second language learners into extended spoken exchanges marketed as conversational practice. Drawing on conversation analysis, this study asks what kind of interactional environment such applications provide for learners bringing human interactional competence to a machine partner. Analysis of eighty-four screen-recorded sessions between six Korean EFL learners and the application Speak documents three patterns. Where a learner’s response displays a misunderstanding of the prior question, no third-position repair occurs; the response is ratified and built upon. Where automatic speech recognition registers a turn as different words than the learner produced, learners repair with variable success or let the divergence pass. In an extended case, repeated repair fails, and the learner states explicitly that shared understanding cannot be reached. The application’s turn design allocates no sequential position for repair-relevant work, displacing that labor onto the learner. We argue that what these applications afford is therefore closer to monologic production than to conversational practice, and claims made for such tools in language curricula should be calibrated accordingly.