When ChatGPT gets it wrong: investigating third position repair by L2 learners at varying proficiency levels

This study explores how 30 Korean English as a foreign language (EFL) learners at three proficiency levels (CEFR A2, B1, and B2) manage third position repair during voice-based role-play tasks with ChatGPT-4. Over 12 hours and 30 minutes of recorded interactions were analyzed using conversation analysis, complemented by retrospective verbal reports (RVRs) to capture both interactional patterns and learner perceptions. The findings show clear proficiency-related differences. A2 learners ( n = 10) consistently abandoned repair, either by switching to their first language or abruptly terminating the conversation, and none of their TPR sequences resulted in successful repair. B1 learners ( n = 10) managed repair partially through simple repetition or confirmation checks but often prioritized progressivity over intersubjectivity; approximately 50% of their TPR sequences were successful. B2 learners ( n = 10) consistently produced TPR and maintained intersubjectivity through a wider range of strategies, achieving an 80% success rate. RVRs show that limited repair knowledge and ChatGPT’s technical constraints contributed to frustration and demotivation, particularly among lower-proficiency learners. Overall, the findings suggest that ChatGPT’s technical features can function as both constraints and learning opportunities, underscoring the enduring value of EFL teachers in mediating learner experiences and optimizing the pedagogical potential of AI chatbots.