Exploring Generative AI as a Roleplay Interlocutor in L2 Task‐Based Pragmatics Learning: Comparing ChatGPT‐Learner and Technology‐Mediated Peer Interactions

Abstract This study compared task‐based interaction with ChatGPT and peer interaction in the effectiveness of learning L2 speech acts (requests and refusals) and explored L2 learners' perceptions towards task performance. Forty Korean EFL learners were randomly assigned to either the peer interaction group ( n = 20) or the ChatGPT group ( n = 20). Both groups participated in a pretest, two tasks, a posttest, a delayed posttest, a post‐task survey, and stimulated recall. During the scenario‐based roleplay tasks with either a peer or ChatGPT, the target speech acts were elicited, and both participants and ChatGPT were required to provide feedback on the appropriateness of the speech acts. The overall appropriateness ratings of the target speech acts were analyzed and compared across the groups. The amount and types of pragmatic‐related episodes (PREs) that occurred during tasks were also analyzed and compared across the groups. Findings revealed that for requests, ChatGPT‐based interactions led to a higher degree of short‐term learning, whereas peer interaction led to better long‐term retention. For refusals, only the ChatGPT group showed learning gains. The amount of exposure to appropriate forms during PREs was significantly higher in the ChatGPT group than in the peer interaction group. Survey and stimulated recall showed an overall positive experience of task performance with ChatGPT.