Abstract
INTRODUCTION: Peer role-play is a widely used pedagogical strategy for fostering intercultural communicative competence (ICC), yet its effectiveness is often limited by the lack of authentic cultural context and the prevalence of foreign language speaking anxiety. This study, therefore, examines the potential of voice-based generative AI (GenAI) as a scalable, high-fidelity alternative. METHOD: Using a quasi-experimental design, two intact classes of undergraduate EFL learners in China were assigned to either a GenAI-partnered experimental group (N = 29) or a peer-partnered control group (N = 30) during a 10-week ICC course. A mixed-methods approach was employed to analyze questionnaires, learning performance, chat logs, and interviews, drawing on t tests, ANCOVA, epistemic network analysis (ENA), and thematic analysis. RESULTS AND DISCUSSION: Quantitative results showed that the experimental group significantly outperformed the control group in both intercultural speaking performance and perceived ICC. Learners interacting with GenAI also reported a significantly greater reduction in speaking anxiety. ENA revealed a cohesive socio-cognitive triangle in learners' engagement patterns, characterized by the integration of social politeness, cognitive repair, and proactive topic extension. Qualitative findings further indicated that learners experienced the GenAI environment as a psychologically safe space that encouraged risk-taking and as a cultural informant that provided just-in-time pragmatic feedback, despite perceived limitations related to missing paralinguistic cues. Overall, the study suggests that GenAI interlocutors can effectively simulate a culturally distinct "other," offering a psychologically safe and cognitively engaging pathway for ICC development. It argues for the systematic incorporation of AI-scaffolded intercultural practice alongside, and in some cases in place of, traditional peer-based activities in EFL and L2 curricula.