Beyond Cognitive Offloading: A Process–Agency Framework for Distinguishing AI Scaffolding from Cognitive Outsourcing in Higher Education

Generative artificial intelligence (GenAI) can improve task performance while changing the cognitive activity through which learning occurs. Yet prevailing contrasts between beneficial assistance and harmful dependence often conflate what is delegated, when support enters the task, how the learner responds, and who controls substantive epistemic decisions. This conceptual article develops a process–agency framework for analysing discrete GenAI-use episodes in higher education. Rather than treating cognitive offloading, scaffolding, and outsourcing as equivalent taxonomic categories, it locates them within a hierarchical decision structure. Cognitive offloading is the broad externalisation of cognitive work; educationally efficient offloading occurs when peripheral or mastered operations are externalised while focal work remains with the learner. When AI participates in a learning-critical operation, a scaffold-oriented design becomes enacted AI-supported scaffolding only if the learner generates, evaluates, revises, and progressively reclaims responsibility. Cognitive outsourcing occurs when AI retains primary generative control over the learning-critical operation before meaningful learner engagement. Productive struggle is treated as a quality of enacted learning activity, not a moral requirement or separate classification criterion. The framework separates ex ante conditions, enacted functional profile, in-episode epistemic control, and post-support outcomes. A set of falsifiable propositions specifies how learning-objective criticality, timing, generative control, prior knowledge, regulatory capacity, AI literacy, and assessment design may shape assisted performance and independent learning.