Custom GPT as mediator: Dynamic Assessment with beginner KFL learners

Abstract

This study examines how a custom GPT-based chatbot can mediate learner development within the Zone of Proximal Development (ZPD) through dynamic assessment (DA) with beginner-level learners of Korean as a Foreign Language. The model was fine-tuned using OpenAI’s My GPT platform, with a custom prompt specifying graduated mediation, responsive behavior guidelines, and target grammar points. Specifically, the study investigates how GPT operationalizes scaffolding processes in text-based dialogue by sustaining interaction, providing form-focused feedback, and adjusting support contingent on learner responsiveness. Ten English-speaking students in a Korean course at a U.S. university interacted with the chatbot weekly over four weeks. Qualitative analysis of 280 learner-GPT turns identified three mediation types: conversational, instructional, and developmental. Through these, the chatbot maintained natural and level-appropriate dialogue, delivered graduated mediation aligned with learner responsiveness, and used accurate learner responses as springboards to guide movement from the Zone of Actual Development toward the ZPD. Complementary quantitative measures showed higher uptake rates and significant gains in mean length of sentence and lexical diversity. These findings suggest that large language models, when carefully tuned, can emulate core principles of Vygotskian mediation and foster human-AI co-construction of learning within scaffolded interaction.