Feedback and Reconstruction in Learning with Generative AI: An Integrative Review and Proposed Framework for Unaided Performance

Abstract Generative artificial intelligence (GenAI) may help students produce higher-quality work; however, its effects on their future ability to work independently remain unclear. This integrative review examines this distinction by analysing research on feedback, metacognitive monitoring, cognitive offloading, and reconstructive learning activities. Targeted web searches and primary-source tracing informed a focused synthesis of adult learning studies and relevant foundational research. Three features may help explain variation in findings across studies: the information feedback conveys, the reasoning learners are expected to undertake and demonstrate, and the conditions under which learning is assessed. The proposed framework distinguishes between a regulatory pathway, in which feedback informs judgments and learning choices, and a practice pathway, in which reconstruction prompts relevant reasoning. This distinction helps explain why learners may become better calibrated without learning more, or learn more without becoming better calibrated. Feedback is expected to enhance the benefits of reconstruction when it improves engagement or reasoning quality, although its incremental benefit may be limited when the activity already provides comparable practice. Four conditional propositions describe these relationships and suggest comparisons that could help distinguish them from the effects of additional instruction, exposure, time, and motivation. The framework provides a foundation for testing when feedback and reconstruction support performance in the absence of assistance. The available evidence is not yet sufficient to establish the full causal pathways or lasting changes in capability, underscoring the value of delayed assessments, transfer tasks, and longitudinal research.