When AI Enters the Pencil Box: A Longitudinal Study of Child–LLM Writing Collaboration in a Chinese Primary School

We report a qualitative study spanning more than six months with upper-primary students (ages 9–12) in southern China using a local LLM (Doubao) for writing. Two interview waves with ten students (nine re-interviewed) and six teachers (three re-interviewed) reveal a set of recurring shifts in our focal cohort: from novelty to strategy, from anthropomorphic “companion/virtual teacher” to tool-like framing, and from quick unsticking to guarded use shaped by dependence, fairness, and exam-transfer Footnote1 concerns. Students valued speed and breadth but noted mechanical tone, age/style mismatch, and weak transfer to AI-free tests. We map a Control × Dependency framework and before/after transitions, and surface tensions around classroom norms and equity amid policy ambiguity and uneven AI literacy. We suggest process-visible, agency-preserving integration: sentence- and phrase-level provenance with mandatory micro-edits, task-specific scaffold ceilingsFootnote2, and process-anchored assessment schemes that explicitly factor in AI-use traces.