Abstract

This eight-week randomized study examined how the order of generative artificial intelligence (GenAI) feedback shaped revision and near-term independent writing among 132 Chinese English-as-a-foreign-language (EFL) undergraduates. Students received argument-first (AF; n = 67) or language-first (LF; n = 65) feedback; 64 pairs underwent within-class, within-pair randomization, and four unmatched students were randomized within class. Both modules were independently generated from each initial draft (D0) and locked before release. Students completed two revisions (D1 and D2) across three tasks and a new-topic, AI-free transfer task. Adjusted AF–LF differences in D0–D2 change favored AF for overall quality (0.811 points, 95% CI [0.424, 1.198]) and argumentation (1.102 points, 95% CI [0.668, 1.536]), but not language quality (0.020 points, 95% CI [−0.381, 0.420]). Revision depth and implementation showed stage-specific crossovers. Language contrasts did not survive primary Holm adjustment, although complete-pair sensitivity analysis detected a crossover. Argumentation transfer favored AF (1.057 points, 95% CI [0.508, 1.606]); overall transfer was less consistent, with the complete-pair interval including zero (0.685 points, 95% CI [−0.111, 1.481]). Module type was structurally linked to release stage, so stage-specific contrasts describe sequence-by-stage configurations rather than isolated content or order effects. The findings support cautious use of staged feedback when argumentative development is the instructional priority.

This work is licensed under a Creative Commons Attribution 4.0 License.