Layer-sensitive cognitive offloading in generative AI-assisted writing: supported performance and independent no-AI outcomes.

INTRODUCTION: Writing with generative artificial intelligence (GenAI) can improve a supported product without necessarily strengthening the competence that remains after support is removed. This study examined that distinction through a layer-sensitive account of cognitive offloading. METHODS: In this eight-week quasi-experimental classroom study, 180 Chinese undergraduates from six intact English academic-writing classes entered the study and 168 completed the protocol. Two classes were assigned to each of three conditions: no-AI writing, bounded AI support with compulsory reflection, and open AI collaboration. Students completed a baseline task, three intervention assignments, and a supervised Week 8 independent no-AI near-transfer task. RESULTS: Open AI collaboration had the highest observed supported-writing mean (M = 4.02), although the class-clustered open-bounded contrast was imprecise (adjusted difference = 0.21, 95% CI [0.06, 0.36], wild-cluster p = 0.064). At Week 8, the bounded-support condition showed higher adjusted outcomes than open collaboration in writing quality (difference = 0.27), higher-order thinking (0.35), argument depth (0.42), and independent revision quality (0.39); exact wild-cluster p values ranged from 0.050 to 0.063. Within the 112 AI-exposed students, open collaboration involved more prompts, greater text incorporation, less self-generated text, and deeper idea and reasoning offloading. An adjusted associative decomposition showed that open collaboration predicted higher aggregate offloading (a = 0.77, SE = 0.06), which was associated with lower Week 8 higher-order thinking (b = -0.45, SE = 0.08); the bootstrap indirect estimate was -0.34, 95% CI [-0.48, -0.21]. Self-regulated writing attenuated, but did not eliminate, this negative association. DISCUSSION: Because conditions were allocated by only six intact classes and the bounded condition combined delegation restrictions with compulsory reflection, the results are interpreted as classroom-level, mechanism-consistent associations rather than definitive causal effects. The findings distinguish performance achieved with GenAI from competence demonstrated independently after support is removed.