GENERATIVE AI AND CHILDREN'S LEARNING: WHAT CURRENT EVIDENCE CAN AND CANNOT TELL US: An Evidence-Informed Perspective on Cognitive Offloading, Learning Processes, and Research Priorities

Generative artificial intelligence (AI) is entering children’s environments, but evidence about its developmental implications remains uneven. This perspective asks what current research can and cannot establish about generative AI and learning in middle childhood (ages 6–12). Rather than proving inherent harm or universal benefit, current studies show AI changes supported task performance and instructional design matters, while direct evidence regarding delayed unaided learning in children remains limited. The paper brings together research on children’s use of generative AI, cognitive offloading, epistemic trust, instructional design, and AI-assisted education. It distinguishes cognitive scaffolding, which preserves the learner’s engagement in essential cognitive operations, from cognitive substitution, where the system performs those operations. This distinction serves as an organising lens. Supported performance should not be mistaken for learning: delayed unaided retention, transfer, and claim evaluation must be measured directly. While recent usage studies show AI is present in children's lives, usage prevalence does not establish developmental outcomes. Positive tutoring experiments with older students, mixed findings from adults, and emerging child-focused research indicate effects depend heavily on the design of assistance, the learner, and the task. The paper concludes with testable predictions and a staged research agenda, beginning with studies that validate task-level measures before carefully designed research with minors.