Multimodal AI avatars in L2 speaking: A time-lagged mediation analysis

This study examines multimodal AI-supported speaking instruction in relation to EFL learners' speaking performance, self-efficacy, and willingness to communicate (WTC), drawing on Social Cognitive Theory (SCT). Despite growing evidence of generative AI in language learning, there is limited evidence on how AI-supported instruction relates to learners' psychological processes over time. Using an 18-week quasi-experimental, time-lagged design, the study involved 380 undergraduates. The experimental group ( n = 180) received multimodal AI-supported TBLT speaking instruction. The control group ( n = 200) received regular TBLT-based speaking instruction. Semi-structured interviews were also conducted to explore students’ learning experiences. Repeated-measures ANOVAs showed greater gains in speaking performance, self-efficacy, and WTC in the AI-supported group. PLS-SEM supported a partial mediation model. Instructional condition was associated with mid-intervention self-efficacy (T2), which was subsequently associated with WTC and speaking performance at T3. Interview findings indicated generally positive learning experiences while also revealing challenges in AI-supported interaction. These findings suggest that multimodal AI-supported speaking instruction may be associated with gains in speaking performance and WTC partly through self-efficacy.