Cognitive Offloading in the Age of Generative AI: Effects on Memory, Learning, and Executive Function - A Narrative Literature Review

Synthesis: Cognitive offloading — handing mental work to an external aid — used to mean looking up a fact or reaching for a calculator. Generative AI moves the boundary: a model can now do the reasoning, synthesis and evaluation that learning depends on. Papasideri and colleagues review a decade of work on what that shift costs. Their answer is consistent across studies: offloaded material is remembered less well, self-regulated effort thins out, and the effects are worst when the tool is unstructured and the task is complex. The same review finds that the cost is not fixed — tasks that demand justification, or that withhold the tool at the point of assessment, keep the learner in the work.

Key Findings

1. Offloading now reaches the processes that produce learning. Models absorb reasoning, synthesis and evaluation, not just recall, so the thing being skipped is the thing that builds understanding.
2. Supplied material is remembered less durably. Several studies extend the "Google effect": content a model hands over is encoded worse than content the learner produces or retrieves.
3. Performance can improve while the learning process weakens. Short-term output rises while self-monitoring, planning and effort fall — a dissociation that recurs across the evidence.
4. Higher-order tasks carry the most risk. Gains are more reliable for recall and comprehension than for analysis, evaluation and creation, where offloading does the most damage.
5. Guidance changes the outcome. Instructor-guided or scaffolded use, and prompts that require justification, reduce the cost; unstructured use increases it.
6. Younger and more dependent users are more vulnerable. Surveys find the negative association with critical thinking strongest among students who use the tool most.

What the review covers, and how

This is a narrative literature review, not a systematic one: no record flow, no pooled effect sizes, no new data. Studies were gathered through targeted searches on offloading, generative AI, memory, self-regulation, critical thinking and executive function, with reference lists hand-searched, and most primary work appeared between 2023 and 2026. Inclusion required a measured outcome tied to model use, or a rigorous synthesis that contextualized such findings; work on only economic, ethical or technical questions was excluded.

Why generative AI is different

Older offloading research studied notebooks, calculators and search, and found a trade: freeing resources for higher-order work while leaving the offloaded representation unpracticed. Two traditions frame it — the extended-mind thesis, which treats tools as functional parts of cognition, and the metacognitive model, which treats offloading as a cost-benefit decision. Generative AI strains both, because a model that writes the argument also performs the work the learner was supposed to do. The review therefore separates "cognitive surrender" — adopting an output with minimal scrutiny — from dependent offloading, which displaces deep processing, and autonomous offloading, which supports engagement. The studies that show harm are mostly about the first two.

What the studies found

The clearest evidence concerns memory. In a cluster-randomized study of senior high school students, structured model use raised both self-regulated learning scores and offloading, and offloading was negatively associated with performance on a two-week delayed retention test; once offloading was accounted for, the direct AI effect on retention was not reliable — which points at offloading as the mechanism. A mathematics experiment found assistance cut problem-solving time and the knowledge built from the problems, and the review adopts "metacognitive laziness" for the habitual avoidance of monitoring it observes elsewhere. A meta-analysis of higher-education outcomes found positive overall effects, smaller for higher-order skills and larger when use was instructor-guided.

The most direct physiological evidence is an EEG study comparing essay writing unaided, with search and with a model: connectivity was strongest unaided and weakest with the model, and participants who returned to unaided writing later showed under-engagement. The review treats that study as provisional given its sample size. Surveys add a correlation between frequency of use and self-assessed critical thinking, mediated by offloading and strongest among younger users, and experiments on persistence find that access to help reduces willingness to keep working without it.

Four boundary conditions

Four moderators recur. Instructional structure helps: guided use with planning, monitoring and evaluation produces better learning than unstructured use. Task complexity matters: benefits are reliable for recall and comprehension, risk is highest for analysis and creation. Mode matters: autonomous use preserves engagement while dependent use and surrender track with weaker retention. And user characteristics matter, with younger and less experienced users more vulnerable.

The authors' conclusion is a qualified offloading-cost hypothesis. A metacognitive-control account fits the evidence better than technological determinism: active engagement with the output, not use itself, decides the consequence. They also suggest the umbrella term may now be too coarse, and that the practical question is under what conditions of task design and guidance a model supports rather than supplants learning.

What this means for practice

• Instructors: require an unaided attempt first — a draft, an estimate, an explanation — before the model is opened. Without one, there is nothing for the student to compare against.
• Instructional designers: build justification into the task. Prompts that ask learners to defend or critique an answer reduced offloading and improved argument quality in the studies reviewed.
• Assessment designers: grade the process. Drafts, reasoning steps and verification carry what the polished artifact cannot, and delayed retrieval tasks expose what offloading left unpracticed.
• Administrators and curriculum leads: schedule AI-free practice rather than assuming students will choose it. The memory findings point to retrieval without the tool as the mechanism that repairs the gap.

Limitations

• Several key studies rely on self-report measures of critical thinking and offloading tendency, which are open to social-desirability and introspection bias.
• Neurophysiological work uses small samples, limiting power and generalization.
• Most designs are cross-sectional, survey-based or short-duration experiments, which rules out strong causal claims about long-term cognitive change.
• Operational definitions of offloading, dependency and metacognitive engagement vary across studies, which prevents pooling.
• The models studied evolve quickly, so findings from earlier generations may not transfer to current tools.
• As a narrative review it offers no record flow or pooled effects, and its evidence skews toward higher-income educational and knowledge-work settings.

Connected Concepts

• Cognitive Offloading — the construct the review extends to generative AI
• Executive Function — the third outcome domain, alongside memory and self-regulation
• Self-Regulated Learning — the processes that thin out under heavy reliance
• Critical Thinking — the self-assessed outcome most often reported
• Metacognition — monitoring and regulation as the moderators that decide the cost

Connected Articles

• Effects of LLM use and note-taking on reading comprehension and memory: A randomised experiment in secondary schools — an experiment where the chatbot condition produced less learning than writing
• Individual Differences in Reliance on Generative AI in Schoolwork: The Role of Executive Functioning Difficulties, Perceived Usefulness, and Habitual Use — individual differences in reliance and executive function
• Academic Stress Mediates the Link Between ChatGPT Dependency and Executive Dysfunction — dependency, stress and executive dysfunction in a mediation model

Citation

Papasideri, V., Chatziioannou, S. S., & Palaiologos, P. (2026). Cognitive Offloading in the Age of Generative AI: Effects on Memory, Learning, and Executive Function - A Narrative Literature Review. Journal of Clinical Psychology and Mental Health, 4(1), 1-6. https://doi.org/10.59657/2993-0227.brs.26.041