Generative AI and adaptive systems for customising help-seeking scaffolds: A systematic review

Abstract

Help-seeking is crucial in self-regulated learning (SRL), but generic scaffolds often do not meet diverse learner needs. This review examines AI, including large language models (LLMs), in customising help-seeking scaffolds, their effects, and methodological and ethical constraints. Using SRL theory, the review analysed empirical studies on AI-driven help-seeking support, focusing on evidence, impacts versus standard scaffolds, evaluation methods, and gaps. The original searches in IEEE Xplore, Web of Science, ERIC, and ScienceDirect on 28 May 2025 identified 3627 records, of which 3576 remained after deduplication; multi-reviewer screening in Rayyan using PRISMA procedures identified 31 studies; a supplementary search of Scopus and the ACM Digital Library (9 July 2026), restricted to the original review period, added 17 further studies, yielding 48 studies for narrative synthesis. Title-and-abstract screening showed high raw agreement (87.3%) but low chance-corrected agreement (Cohen’s κ ≈ 0.05; Fleiss’ κ ≈ 0.17), resolved through structured adjudication. Interventions (2017–2025) included rule-based tutoring, learning-analytics dashboards, LLM-based textbooks, writing feedback, and coding assistants, providing conversational, intent-sensitive scaffolding. Most studies showed improved performance and engagement, especially for lower-performing learners, though null or mixed effects and inconsistent satisfaction were common. Limitations included sample scope, duration, measurement variability, usability, and risks of hallucinations, bias, privacy, equity, and overreliance. The review concludes AI-customised help-seeking scaffolds are promising but context-dependent, highlighting methodological and design priorities for SRL-aligned, responsible AI support.