Answers are grounded in this paper and cite their sources.
Although human-AI conversational tools are being adopted at an increasing rate for second language (L2) oral practice, how such interactions are psychologically associated with speaking anxiety and willingness to communicate across different settings remains poorly understood. We examined whether, in a moderated mediation framework, human-AI conversational experience (HAICE) is associated with L2 speaking willingness to communicate (SWTC) through L2 speaking self-efficacy (SSE), and whether L2 speaking anxiety (SA) moderates the SSE-to-SWTC path. Survey responses were gathered from 486 Chinese and 412 Malaysian university students after completing a standardized AI-mediated English conversation task. We employed confirmatory factor analysis, structural equation modeling, bootstrap mediation tests, latent interaction modeling and multi-group invariance analysis. HAICE was positively associated with both SSE (β = 0.34, p < 0.001) and SWTC (β = 0.10, p = 0.019). The pattern of associations was consistent with partial mediation of the HAICE–SWTC link through SSE (indirect effect = 0.13, 95% CI [0.08, 0.18]), accounting for 56.5% of the total association. SA was negatively associated with SWTC (β = −0.19, p < 0.001), and a significant SA × SSE interaction (β = 0.14, p = 0.003) was consistent with a buffering pattern. The overall model explained approximately 27% of the variance in SWTC ( Δ R interaction 2 = 0 . 018 ). Configural, metric and partial scalar invariance across the two national samples were supported. The buffering interaction was significant in both the Chinese (β = 0.17, p = 0.002) and Malaysian (β = 0.11, p = 0.031) samples; a formal Wald test did not indicate a statistically significant between-group difference (χ 2 (1) = 1.87, p = 0.171). Qualitative responses characterized AI conversation as a perceived “low-threat rehearsal space.” These results broaden the L2 willingness to communicate literature by showing that AI-mediated conversation is associated with speaking willingness through a dual cognitive-affective pathway and that self-efficacy attenuation of speaking anxiety may vary across national contexts with different sociolinguistic profiles. Because the design is cross-sectional, all associations should be interpreted as correlational, and longitudinal and experimental replication will be needed to establish temporal precedence and causal claims.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: