From chatbot to language-learning partner: learner–LLM configuration, affordances, and regulatory demands in foreign language learning

Large Language Models (LLMs) are increasingly used as flexible resources for self-directed foreign-language learning, yet limited research explains how learners configure these systems as pedagogical partners. This study examined pedagogical role construction, modality–pedagogy alignment, affordance realization, breakdowns, and regulatory demands in learner–LLM interaction. Using a retrospective observational computational content-analysis design, 117 first-person public posts and comments, representing 80 pseudonymous participants, were analyzed from seven public forums and X. Structured multi-label coding and quantitative analyses identified eight pedagogical roles and examined relationships among interaction modalities, learning activities, perceived affordances, and breakdowns. Results showed that learners reported configuring LLMs as multifunctional pedagogical partners, most prominently as interlocutors, while 65.8% of accounts involved multiple roles. Voice interaction was significantly associated with conversation and pronunciation, whereas text was associated with vocabulary learning, supporting modality–pedagogy alignment. More orchestrated learning configurations were associated with greater perceived benefits, particularly personalization and convenience, although an apparent difference in activity breadth was substantially attenuated once the activities used to define orchestration were excluded from the count. However, 43.6% of accounts reported concerns, revealing function-specific breakdowns involving accuracy/language reliability, access/privacy/system constraints, pedagogical sustainability, interaction/continuity, and voice/technical reliability. The findings support an empirically informed framework conceptualizing learner–LLM interaction as an AI-mediated regulatory process involving role construction, modality and activity configuration, affordance realization, monitoring, breakdown detection, and regulatory demand. Psychological self-regulation and objective proficiency gains were not directly measured.