Metacognitive Filtering and Cognitive Offloading in AI-Assisted L2 Writing: A PRISMA Guided Process-Tracing Synthesis

This study conducted a PRISMA-guided systematic review of 33 empirical studies examining learner interactions with AI-mediated feedback in L2 writing between 2010 and 2025. Evidence was synthesized from process-tracing methods, including keystroke logging, eye-tracking, screen capture, interaction logs, and draft-history analytics. The findings revealed two recurring interaction profiles. Higher-regulation learners typically engaged in selective uptake, recursive evaluation, and extended processing of AI feedback, whereas lower-regulation learners more frequently indicate rapid acceptance and reduced evaluative engagement. Across studies, micro-level feedback was associated with shorter revision latencies and burst-editing patterns, whereas macro-level feedback was associated with longer processing times and repeated revision cycles. The review also identified substantial methodological variability in trace precision, AI transparency, multimodal triangulation, and reproducibility practices. Overall, the evidence suggests that the effectiveness of AI-mediated feedback depends not only on the quality of the feedback provided but also on how learners regulate and engage with it during revision. Greater methodological transparency and more rigorous process-tracing designs are needed to strengthen future research on AI-supported writing development.