A mixed-methods study on the contributions of AI-supported mediation modes to L2 learners’ speaking skills

Purpose Framed within sociocultural theory, this study examined the comparative contributions of artificial intelligence (AI) mediation, human mediation, combined human–AI mediation, and no-mediation conditions to the speaking development of Iranian English as a foreign language (EFL) learners. It also explored learners' experiences with and perceptions of these different mediational modes.Design/methodology/approach Adopting a sequential explanatory mixed-methods design, the study involved 128 intermediate-level Iranian EFL learners randomly assigned to four conditions: AI mediation, human mediation, combined human–AI mediation, and no mediation. Speaking performance was assessed before and after instruction using the IELTS Speaking Band Descriptors. Qualitative data were collected through guided reflective narratives and semi-structured interviews and analyzed thematically.Findings The combined human–AI mediation group achieved the highest adjusted posttest mean, followed by the AI mediation, human mediation, and no-mediation groups. Thematic analysis identified two overarching themes in learners' experiences and perceptions: (1) perceived reciprocity and suitability of mediation and (2) experienced affective safety in speaking. AI mediation was generally perceived as fast and low-pressure, whereas human mediation was valued for its contextual responsiveness and emotional sensitivity.Originality/value This study extends sociocultural perspectives on AI-assisted language learning by comparing the contributions of AI, human, and combined human–AI mediation to EFL speaking development. It highlights the potential value of integrating the efficiency and low-pressure environment of AI with the contextual and affective sensitivity of human support, suggesting that human–AI co-mediation may provide complementary forms of assistance responsive to learners' linguistic and affective needs.