Beyond the Glossary Trap: Register-Sensitive Evidence Practices in DDL-With-AI Mediation Training

Abstract Data-driven learning (DDL) has long positioned learners as investigators of authentic language data. With the rapid integration of generative artificial intelligence (GenAI) into language classrooms, however, little is known about how learners interactionally negotiate corpus evidence and AI-generated reformulations in practice. This study examines how DDL is enacted when combined with large language models (LLMs) in a multilingual Master’s course in Mediation English at an Italian university. Across three instructional cycles (business, healthcare, legal), students produced human renditions, generated AI-assisted reformulations using freely accessible LLMs, and evaluated alternatives through corpus consultation. The design evolved from relatively autonomous corpus querying ( hard DDL) to scaffolded interpretation of curated extracts ( soft DDL), allowing comparison of interactional practices under varying degrees of teacher mediation. Analysis draws on written artefacts (translations, prompts, corpus queries, justificatory commentary) analysed using five descriptive analytic lenses (problem orientation, query type, evidential reasoning, AI positioning, and evidence-decision alignment). Findings show a progressive shift from frequency-driven lexical selection and default AI adoption toward register-sensitive reasoning and evidence-mediated AI positioning. Rather than functioning as a shortcut to fluency, AI became a contestable epistemic interlocutor when embedded within structured corpus comparison. The study contributes classroom-based evidence to discussions of DDL with AI by conceptualising corpus consultation as a mediating practice that redistributes authority between human judgment and machine-generated language within digitally augmented learning ecologies.