Abstract
This study examines how discussion expressions are learned and evaluates the potential and limitations of AI-based practice. Expression frequencies in final test discussions by advanced international students were compared with those in simulated discussions involving two humans and one AI. Some expressions practiced in class appeared frequently in tests but rarely in AI interactions, while others were overused by the AI. Although the AI supported discussion organization and summarization, it struggled to express hesitation, opposition, and backchannels. These imbalances may influence how humans perceive the interaction, suggesting that AI is useful for preparation but real human discussion remains essential.
https://ror.org/04bcbax71 
