Evaluating AI-Enhanced Collaboration in EFL Academic Writing: A Longitudinal Mixed-Methods Study

This longitudinal mixed-methods investigation addressed the unresolved question of how different modes of AI-enhanced collaboration shape EFL academic writing across content enrichment, linguistic accuracy, and organizational coherence. A four-arm, stratified, parallel-group randomized controlled trial (Randomized Controlled Trial [RCT]) with concealed, centralized allocation and assessor blinding was implemented over an 8-week hybrid program (2 h in-person +2 h online weekly; 32 contact hours). Randomization used permuted blocks (size = 8; 1:1:1:1) within university, baseline proficiency tertiles, and gender strata; analyses followed intention-to-treat with multiple imputation (m = 50; Rubin’s rules) and prespecified sensitivity checks (complete-case, per-protocol). Quantitatively, outcomes were captured via rubric-based assessments, automated writing analytics, and AI engagement indices; qualitatively, semi-structured interviews (n = 88) and validated surveys were analyzed through reflexive thematic analysis. Integration occurred at design, methods, and interpretation using joint displays and triangulation. The content-enrichment intervention produced the largest and most durable gains (e.g., d > 0.80), including a substantial adjusted mean difference in overall writing versus control (Δ = 1.95, p < .001). Multivariate analyses (MANCOVA) showed a significant group effect after adjusting for pretest scores (p < .001). Path analysis indicated that perceived AI utility and digital literacy significantly mediated performance gains. Thematically, AI functioned as a collaborative partner, though dense feedback sometimes elevated cognitive load.