Generative AI, Cognitive Offloading, and Learner Agency in Higher Education: A Scoping Review

Synthesis: Wang, Wang, Yang and Ren conducted a PRISMA-ScR-informed scoping review of research on generative AI, Cognitive Offloading and learner agency in higher education, searching six database sources on 24 June 2026 for English-language journal literature published from 2022 onward. From 8020 retrieved records, 2693 duplicates were removed, 5327 records were screened by title and abstract, 691 full texts were assessed, and 123 studies entered the final charting matrix and synthesis. The review maps a dual-pattern account: agency-supportive patterns run through Self-Regulated Learning, Self-Efficacy, Feedback Literacy and reflective engagement, while agency-eroding patterns run through cognitive offloading, overreliance, dependence and uncritical uptake. It reports no pooled effect — the synthesis is configurative and interpretive, the corpus is heterogeneous, and no formal critical appraisal or risk-of-bias assessment was conducted.

Key Findings

• The corpus is recent and concentrated in a few sources. Of the 123 included studies, 5 were published in 2024, 34 in 2025 and 84 in 2026. After deduplication the retained sources were Scopus (67), EBSCOhost Education Source (25), APA PsycInfo via ProQuest (15), EBSCOhost ERIC (8), ProQuest Education Collection via ProQuest Social Science Premium Collection (5), EBSCOhost combined/exported source (2) and Web of Science Core Collection (1).
• Evidence types were mostly empirical but methodologically mixed. The corpus held 115 empirical studies, 7 review/evidence-synthesis articles and 1 conceptual/framework source. Method labels included experimental/intervention/quasi-experimental (33), mixed-methods/Q-methodology (25), quantitative survey/correlational (24), qualitative (20), scale development/validation (6) and empirical study not further classifiable (7).
• The literature is concentrated in a few contexts. Country/region signals most often identified China (58) and the United States (46), and 15 studies had no clearly identifiable country. Educational-context signals were most common for research/postgraduate learning (100), discipline-specific higher education (77), language/EFL/ESL (71), writing/feedback (70) and STEM/problem solving (70).
• Learner agency is conceptualized as multidimensional, not inferred from tool use. Agency/autonomy signals appeared in 121 of 123 studies, and 86 studies contributed to the agency/autonomy conceptualization theme family. The review reports frameworks built on key abilities, active actions and essential mental characteristics, and on receptive, resistive, resourceful and reflective forms of engagement.
• Both agency-supportive and agency-eroding patterns were near-universal in the corpus. 118 studies contributed to the agency-supportive theme family and 108 to the agency-eroding/offloading theme family. Mechanism-related constructs (self-regulation, metacognition, critical thinking, self-efficacy, trust, feedback literacy, AI literacy, and cognitive offloading/reliance/dependence) appeared in 116 studies, and boundary conditions/pedagogical responses in 120.
• Screening reliability varied by stage. Title/abstract screening showed Fleiss' kappa = 0.7103 with pairwise Cohen's kappa of 0.7147, 0.5799 and 0.8309. Full-text screening fell to Fleiss' kappa = 0.4884 with pairwise values of 0.3939, 0.3056 and 0.8475, which the authors attribute to the interpretive boundary between broad GenAI-in-education relevance and substantive relevance to agency, offloading or reliance.

How the review was conducted

The search ran on 24 June 2026 across Web of Science Core Collection, Scopus, EBSCOhost ERIC, EBSCOhost Education Source, APA PsycInfo via ProQuest, and ProQuest Education Collection via ProQuest Social Science Premium Collection. Database-specific retrieval was Web of Science 939, Scopus 2524, ERIC 610, Education Source 1316, PsycInfo 726 and ProQuest Education Collection 1905 — 8020 records before deduplication. After 2693 duplicates were removed, 5327 records were screened; 4610 were excluded at title/abstract, 717 reports were sought, 26 could not be retrieved, and 691 full texts were assessed. 568 reports were excluded with reasons — most often general GenAI in higher education with no substantive focus on agency/offloading mechanisms (337) or performance/satisfaction/attitudes only (97) — leaving 123 included studies. The protocol is registered in INPLASY (INPLASY202650152).

Data were charted with a structured extraction matrix and synthesized through descriptive mapping and theory-informed configurative synthesis organized into five theme families. The authors state explicitly that the review does not estimate a single pooled effect and is not a meta-analysis: the framework is an interpretive synthesis of heterogeneous evidence, and recurring mechanisms are described as constructs that appeared repeatedly across the charted corpus rather than as a formally ranked causal hierarchy. No formal critical appraisal or risk-of-bias assessment was conducted.

What this means for practice

• Design for augmentation, not replacement. The review's central practical claim is that the educational value of GenAI depends less on the tool than on pedagogical embedding: scaffolded, guidance-based use (hints, prompts, feedback, intermediate support) was associated with agency-supportive patterns, whereas answer-generating use aligned with efficiency demands was associated with displacement of learner judgment.
• Teach self-regulation and metacognition as conditions, not extras. The literature treats self-regulation, Metacognition and critical evaluation as the conditions under which GenAI becomes educationally productive; Scaffolding such as metacognitive prompts and structured reflection journals was associated with reduced cognitive load, improved performance and self-efficacy, and stronger GenAI literacy.
• Treat AI literacy as evaluative literacy. Students need to judge output quality, recognize risk, formulate productive prompts and use feedback selectively — not merely to operate the tool.
• Do not read confidence as competence. Self-efficacy was ambivalent: more frequent GenAI use was reported to enhance confidence and efficiency while intensifying technological dependence, and dependency was linked to inflated, false self-efficacy, so rising confidence alone does not demonstrate preserved Learner Agency.
• Keep human mediation in the loop. Teacher guidance, dialogic feedback and verification requirements recur as boundary conditions that keep GenAI a cognitive scaffold rather than a cognitive substitute.

Limitations

• The review is limited to English-language peer-reviewed or scholarly journal literature published from 2022 to 24 June 2026, so books, theses, conference proceedings, institutional reports and other gray literature were not captured.
• No formal critical appraisal or risk-of-bias assessment was conducted and the corpus is methodologically heterogeneous, so the conclusions are conceptual and interpretive rather than causal; the authors state that recurring mechanisms have not been established with equal empirical strength across contexts.
• The included literature is concentrated in Language Learning, writing and Chinese/EFL higher education settings (China n = 58; language/EFL/ESL n = 71), which limits generalizability to underrepresented areas such as laboratory-based STEM, design and professional education.
• Many included studies used cross-sectional self-report data, so the temporal dynamics of the dual-pattern account remain untested. The full-text stage also lost 26 of 717 sought reports to retrieval limits, and the GenAI search block emphasized ChatGPT, generative AI, GenAI, large language model, LLM, GPT, GPT-3 and GPT-4 terms, so studies indexed only under other product or conversational-agent labels may have been missed.

Citation

Wang, G., Wang, W., Yang, D., & Ren, J. (2026). Generative AI, Cognitive Offloading, and Learner Agency in Higher Education: A Scoping Review. Behavioral Sciences, 16(7), 1150.