Mediating Between Corpus Evidence and AI Output: Teacher Education for DDL in Hungarian as a Foreign Language

Abstract Data-driven learning (DDL) has become a key approach in language pedagogy, but its application to Hungarian as a foreign language (HFL) is still limited. Unlike teachers of widely taught languages, HFL teachers cannot rely on a wealth of pre-designed resources and often need to create their own materials. The rise of generative AI (GenAI) tools, such as ChatGPT, offers extensive language output but raises concerns about the naturalness and reliability of this data for less-represented languages, such as Hungarian. This paper, drawing on interactional data from a postgraduate teacher education course, examines the processes through which teachers negotiate meaning, epistemic authority, and pedagogical decisions as they work with Sketch Engine and ChatGPT. Focusing on collaborative worksheet design for young learners of Hungarian as a foreign language, the analysis draws on recorded peer interactions, teacher–GenAI chat logs, and post-task reflective accounts. Rather than treating corpus data or GenAI output as authoritative sources, teachers orient to both as provisional resources that require interpretation, validation, and pedagogical mediation. The findings show that doing DDL involves recurrent interactional practices, such as epistemic hedging, peer scaffolding, prompt refinement, corpus consultation, and managing productive breakdowns when GenAI output conflicts with corpus evidence or pedagogical expectations. By analysing how teachers negotiate the relationship between corpus evidence, AI-generated language, and anticipated learner needs, the study highlights DDL as a socially organised, collaborative practice rather than a tool-driven procedure. The findings offer implications for teacher education and for the responsible integration of DDL in resource-scarce language contexts.