Abstract
Introduction: Large language models (LLMs) are increasingly used to provide formative feedback in writing instruction, yet their pedagogical design, feedback quality, and effects on student outcomes require systematic synthesis. Methods: Following PRISMA 2020 guidelines, this review searched Scopus, IEEE Xplore, and ERIC for peer-reviewed studies published between November 2022 and February 2026. After screening and eligibility assessment, 34 studies were included. The synthesis focused on classroom interventions, the quality of LLM-generated feedback, and student learning outcomes. Results: The reviewed studies were conducted across 19 countries, with a strong concentration in university-level writing courses. Most interventions used ChatGPT or GPT-based systems through chatbot interfaces or integrated learning platforms. LLM-generated feedback supported writing productivity, linguistic accuracy, engagement, and feedback literacy, particularly by enabling rapid and iterative revision. However, human feedback remained stronger for higher-order writing issues, including argument development, contextual interpretation, prioritization of revision needs, and dialogic guidance. Discussion: The findings suggest that LLMs are most effective when used as supplementary feedback mechanisms rather than replacements for teachers. Hybrid feedback models that combine AI-generated suggestions with teacher mediation appear especially promising. Future research should strengthen instructional design, feedback literacy, ethical safeguards, and evidence from primary and secondary education contexts.

This publication has 47 references indexed in Scilit: