From Static to Dynamic: Parameter Manipulation as Adaptive DDL Scaffolding in Creative Writing with Generative AI

Abstract Generative AI tools now pervade university classrooms, yet the field lacks empirical understanding of how technical AI configurations translate into adaptive Data-driven Learning (DDL) scaffolding. This gap is crucial in creative writing, where excessive structure can suppress authorship and insufficient support can stall exploration. This study investigates how large language model (LLM) parameter configuration functions as a pedagogical design lever for DDL scaffolding in poetry writing classrooms. Across a three-phase mixed-methods study with 30 university creative writing learners (a 2 × 2 factorial experiment, an offline classroom workshop, and an online classroom workshop), two parameter settings, temperature and top-p, reliably reshaped the distribution of three pre-coded interaction types: Constraint Repair (Type A), Exemplar Giving (Type B), and Surprise Harvest (Type C). Moving from a structured to an exploratory setting, Type C rose roughly seven-fold in Phase 1, with the same direction of effect in the two subsequent classroom phases. Triangulating chat logs, reflection templates, and panel transcripts, Type B was voted most helpful but linked to diminished authorship, while Type C was associated with the highest authorship and satisfaction. We position parameter literacy as a necessary instructor competence for adaptive AI scaffolding.