Answers are grounded in this paper and cite their sources.
Cognitive agency concerns learners’ effortful, self-directed control over reasoning with generative AI (GenAI). Building on an existing implementation-level classification of generative AI use, this review examines how cognitive work is allocated between learners and AI tools. Following PRISMA 2020, searches of Scopus and Web of Science (2021–2025) yielded 93 peer-reviewed studies, appraised with the Mixed Methods Appraisal Tool. A realist synthesis specified an a priori program theory crossing the type of work offloaded (extraneous versus germane; cognitive load theory) with processing of AI output (heuristic acceptance versus analytical scrutiny; dual-process theory), and audited the evidence on each axis under manifest coding (reported evidence only) and latent coding (adding design-based inference). Under manifest coding, 21 studies were located on both axes; all four configurations occurred, two in a single study each, and the nine Cognitive Extension studies reported peripheral uses with analytical scrutiny. Of the 93 studies, 49 reported no learner behavior toward AI output, 69 reported no learning measure independent of AI assistance, and four reported processing evidence together with such a measure. Group differences or part-sample outcomes were documented in 30 studies; one further study reported stage differences, but no learner-, task-, or assessment-level moderator showed an interpretable corpus-level association with substitution. This evidence profile identifies the process and AI-independent learning measures needed to distinguish cognitive extension from cognitive substitution. The review restates the Heterogeneous Pathways of Distributed Agency (HPDA) Model as seven testable propositions and specifies the process-tracing and AI-independent outcome designs needed to test them.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: