Beyond Cognitive Offloading: What Must Remain Internal in Generative AI-Supported Learning?

Generative artificial intelligence (GenAI) can now perform cognitive activities such as explanation, synthesis, generation, evaluation, and reasoning on learners’ behalf. The central educational question is therefore no longer only which tasks can be delegated to GenAI, but which cognitive resources must genuinely remain with the learner after learning. Existing research has addressed this issue through distinctions such as scaffolding versus substitution, different forms of cognitive offloading, the locus of GenAI participation, and task allocation. Building on this work, the present paper shifts the analytical focus to three internal cognitive resources that learners ultimately need to retain: internal knowledge, internal cognitive capability, and internal cognitive control. These correspond to what learners need to know, what they ultimately need to be able to do, and how they need to judge when to use, verify, or reject GenAI. Drawing on cognitive offloading, cognitive load theory, generative learning, the learning-performance distinction, and metacognitive theory, we argue that GenAI may reduce non-target processing and may temporarily perform some cognitive operations, but it should not replace the knowledge, capabilities, or regulatory control that the learning objective itself requires learners to develop. GenAI-supported learning should therefore begin with the learning objective, specify the internal resources that must be retained, determine how and when GenAI should intervene and be faded, and verify learning through retention, transfer, independent performance, and metacognitive calibration. The paper concludes with corresponding design principles and a research agenda, while treating neuroscience as an explanatory basis rather than direct evidence of neural effects caused by GenAI.