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

The rapid diffusion of generative artificial intelligence (GenAI) tools such as ChatGPT into education and knowledge work has revived long-standing questions about cognitive offloading - the practice of delegating mental effort to external aids. Unlike earlier offloading technologies such as calculators or search engines, generative systems can absorb higher-order cognitive work, including reasoning, synthesis, evaluation, and creative generation, raising the possibility of qualitatively new effects on memory, learning, and executive function. This review synthesizes recent theoretical and empirical literature (2016-2026) on cognitive offloading in the GenAI era. Across neurophysiological, behavioral, and survey-based studies, a consistent pattern emerges: GenAI use is associated with reduced neural connectivity during effortful tasks, weaker retention and delayed recall of offloaded material, diminished self-regulated learning and metacognitive engagement, and, in frequent users, lower self-reported critical thinking. These effects are not uniform; they are moderated by task complexity, instructional structure, prior domain knowledge, and how offloading is enacted - with a distinction emerging between “autonomous” offloading that preserves active engagement and “dependent” offloading or “cognitive surrender” that displaces it. The review situates these findings within extended-mind and metacognitive-control theories of offloading, identifies methodological limitations (small samples, cross-sectional designs, self-report reliance), and outlines directions for longitudinal and mechanistic research. Implications for instructional design and AI literacy are discussed.

Cognitive offloading refers to the use of physical or digital action to reduce the internal processing demands of a task-writing a reminder instead of rehearsing it in memory, or using a calculator instead of performing mental arithmetic (Risko & Gilbert, 2016). Offloading is not new, and decades of research on notebooks, calculators, and internet search have shown that it can be adaptive, freeing limited cognitive resources for higher-order work, while also carrying costs for the internal representations that go unpracticed (Sparrow, Liu, & Wegner, 2011; Clark & Chalmers, 1998).

Generative AI changes the scope of what can be offloaded. Where search engines offload the retrieval of facts and calculators offload computation, large language models (LLMs) can offload reasoning, argument construction, evaluation, and creative generation-activities long considered core to higher-order cognition (Noy & Zhang, 2023; Doshi & Hauser, 2024; Woodruff et al., 2024). Because these are precisely the processes that memory research, educational psychology, and executive-function research treat as drivers of durable learning, the arrival of GenAI has prompted renewed and more urgent interest in the cognitive consequences of offloading. Commentators and researchers alike have described a spectrum of concern, from AI-induced “deskilling” and cognitive dependency (Gerlich, 2025; Shukla et al., 2025; Yu et al., 2024) to more optimistic accounts in which GenAI, used deliberately, can scaffold metacognition and self-regulated learning (Tomisu et al., 2025; Yan et al., 2024).

This review addresses three linked questions: (1) How does GenAI use affect memory encoding, retention, and retrieval of offloaded material? (2) What are its effects on learning processes, including self-regulated and metacognitive engagement? (3) How does GenAI use relate to executive function, particularly critical thinking, sustained attention, and independent problem-solving? A fourth, cross-cutting question concerns the boundary conditions under which offloading is adaptive versus detrimental-that is, which patterns of use protect cognitive engagement and which erode it.

Theoretical Framework

Two theoretical traditions anchor this literature. The first is the extended-mind thesis, which holds that cognitive processes can incorporate external tools as functional components of cognition rather than mere aids to it (Clark & Chalmers, 1998). Under this view, offloading to AI is a continuation of a long human practice of distributing cognition across brains, bodies, and artifacts.

The second is the metacognitive model of cognitive offloading proposed by Risko and Gilbert (2016), later formalized as a value-based decision process (Gilbert, 2024). In this account, a person continuously and often implicitly weighs the expected cognitive cost of retaining information internally against the cost and reliability of an external aid, and offloads when the external option is judged more efficient. A well-documented corollary is the “Google effect”: information that people expect to be able to retrieve externally is less well encoded and recalled internally, even though people remember where to find it (Sparrow et al., 2011; Eskritt & Ma, 2014; Kelly & Risko, 2019a, 2019b).

Recent theoretical work argues that generative AI may require a further distinction. Risko and Gilbert's original model concerned the delegation of discrete, well-defined subtasks (e.g., arithmetic to a calculator) while the person retained control of the overall problem structure. Several authors argue that GenAI enables a qualitatively different pattern-termed “cognitive surrender”-in which a person adopts an AI-generated output as their own with minimal scrutiny, relinquishing control of the reasoning process itself rather than delegating a bounded step within it (Shaw & Nave, 2026, as cited in recent replication work). Relatedly, a distinction has been proposed between “dependent” offloading, which displaces deep processing and skill development, and “autonomous” offloading, which uses AI as a scaffold while preserving the user's active engagement with the material (Carter, 2018, as cited in recent GenAI offloading research). This review uses these constructs to organize and interpret the empirical findings below.

This is a narrative literature review rather than a formal systematic review or meta-analysis, reflecting the young and fast-moving state of the empirical base (most primary studies were published between 2023 and 2026). Relevant work was identified through targeted searches of academic databases using combinations of the terms “cognitive offloading,” “generative AI,” “large language models,” “ChatGPT,” “memory,” “self-regulated learning,” “metacognition,” “critical thinking,” and “executive function.” Reference lists of identified papers were hand-searched for additional sources (snowball sampling).

Sources were prioritized for inclusion if they (a) empirically measured a memory, learning, metacognitive, or executive-function outcome in relation to GenAI or LLM use, (b) were peer-reviewed or, where very recent, rigorously conducted preprints from established research groups, or (c) provided theoretical or meta-analytic synthesis of the offloading literature that contextualizes GenAI-specific findings. Work concerned solely with AI's economic, ethical, or technical dimensions, without a cognitive outcome, was excluded. Given the pace of publication in this area, this review should be read as a synthesis of the current state of evidence rather than an exhaustive census; it does not report a PRISMA-style flow of records, and formal effect-size pooling was not attempted given the heterogeneity of designs and outcome measures across the included studies.

Characteristics of Included Studies

Table 1: summarizes the design, sample, and key outcomes of the primary empirical studies, meta-analyses, and theoretical syntheses underpinning the results reported in Section 5. Studies are presented in the approximate order in which they are discussed below, moving from foundational offloading theory through GenAI-specific memory, learning, and executive-function findings.

Effects on Memory

The clearest and most theoretically grounded finding concerns memory for offloaded content. Extending the classic “Google effect” (Sparrow et al., 2011) to generative AI, several studies report that content produced or supplied by an LLM is less durably encoded than content a person generates or retrieves themselves. In a cluster-randomized study of senior high school students, structured GenAI use increased both self-regulated learning strategy scores and cognitive offloading, but offloading was negatively associated with performance on a two-week delayed retention assessment, while the direct effect of AI use on retention was not reliable once offloading was accounted for-suggesting offloading itself, rather than AI exposure per se, is the operative mechanism (IMCC Journal of Science, 2024).

A mathematics-learning study similarly found that GenAI assistance reduced the time students spent studying problems and correspondingly reduced the durable knowledge built from those problems, a pattern the authors distinguish from classical offloading and instead describe as “cognitive surrender” when AI output is adopted with minimal scrutiny (arXiv, 2026, “Faster Completion, Less Learning”). Consistent with this, a review of digital-technology effects on cognition reports convergent behavioral and neural evidence-including changes in prefrontal and anterior cingulate activation-associated with heavy reliance on external digital and AI aids for memory-dependent tasks (Yüce, 2025; Korte, 2020, as cited in recent reviews).

Effects on Learning, Metacognition, and Self-Regulation

A second cluster of findings concerns learning processes rather than memory storage per se. Fan et al. (2025) introduced the concept of “metacognitive laziness” to describe a habitual avoidance of the deliberate cognitive effort normally required to monitor and regulate one's own learning; their study found that GenAI use reduced engagement in key self-regulated learning processes even where short-term task performance improved. This dissociation-better immediate output, weaker underlying process-recurs across the literature and is one of the more consistent patterns identified.

A large-scale meta-analysis of GenAI's impact on learning outcomes in higher education found an overall positive effect of GenAI tool use on learning outcomes, but this effect was significantly moderated by both the cognitive level of the skill being learned and the instructional context: effect sizes were smaller for higher-order skills (analyzing, evaluating, creating) than for lower-order skills (remembering, understanding, applying), and were substantially larger when AI use was instructor-guided rather than left unstructured (CHI EA, 2025). A separate meta-analysis in the Journal of Computer Assisted Learning likewise examined GenAI's effects across cognitive, competency, and affective learning dimensions and identified subject discipline and mode of integration as key moderators of benefit (Zhong, 2025).

Field research on information-seeking behavior similarly finds that conversational GenAI tools tend to reduce overall cognitive effort and metacognitive activity relative to traditional search, and that this reduction in effort is associated with weaker downstream knowledge and skill acquisition, particularly when learners lose access to the tool after having relied on it during initial training (Lee et al., 2025; Bastani et al., 2024; Kumar et al., 2025).

Effects on Executive Function, Attention, and Critical Thinking

The most direct neurophysiological evidence comes from an EEG study conducted at the MIT Media Lab (Kosmyna et al., 2025), which compared essay writing under three conditions: unaided (“brain-only”), search-engine-assisted, and LLM-assisted. Across 54 participants, brain connectivity measured via dynamic directed transfer function analysis was strongest and most distributed in the brain-only group, intermediate in the search-engine group, and weakest in the LLM group. When LLM users were reassigned to the unaided condition in a fourth session, they showed under-engagement relative to those who had practiced unaided from the start, a pattern the authors describe as accumulated “cognitive debt.” The study has drawn methodological critique regarding its sample size, analytic transparency, and reproducibility (commentary in arXiv, 2026), and its findings should be treated as an important but provisional signal rather than a settled result pending replication.

Survey-based work by Gerlich (2025) reported a significant negative correlation between frequency of GenAI use and self-assessed critical thinking ability, statistically mediated by cognitive offloading, with younger and more AI-dependent users showing the lowest scores. A related cross-country randomized experiment by the same author found that this relationship is not fixed: structured prompting protocols that required users to justify or evaluate AI outputs reduced offloading behavior and were associated with deeper reflective engagement and higher argument quality, indicating that offloading is shaped by interaction design rather than being an inevitable consequence of AI use (Gerlich, 2025b, as cited in AI-overdependence review literature).

Complementary experimental work on task persistence found that access to AI assistance reduced participants' willingness to persist on subsequent unaided problems and lowered independent performance once assistance was withdrawn, a pattern the authors interpret as consistent with, but potentially exceeding, classical offloading costs observed with non-AI tools (arXiv, 2026, “AI Assistance Reduces Persistence”). Findings on creativity are more mixed: a systematic meta-analysis found that human-GenAI collaboration produced small but consistent gains in creative output, while simultaneously reducing the diversity of ideas generated, pointing to a potential homogenizing effect on creative and evaluative executive processes even where output quality improves (arXiv, 2025, “Generative AI and Creativity”).

Moderating Factors and Boundary Conditions

Across studies, four moderators recur. First, instructional structure: guided or scaffolded AI use (e.g., prompts requiring planning, monitoring, and evaluation) consistently produces better learning and lower offloading-related costs than unstructured, open-ended use (CHI EA, 2025; IMCC Journal of Science, 2024). Second, task complexity and cognitive level: benefits are more reliable for lower-order tasks (recall, comprehension) than for higher-order tasks (analysis, evaluation, creation), where offloading risk is greatest (CHI EA, 2025). Third, the mode of offloading: autonomous offloading, in which the user remains an active evaluator of AI output, appears to preserve cognitive engagement, whereas dependent offloading or “cognitive surrender,” in which AI output is accepted with minimal scrutiny, is most consistently associated with reduced learning and retention (Frontiers, 2026; arXiv, 2026 “Faster Completion”). Fourth, user characteristics such as age and prior domain expertise moderate dependency, with younger and less experienced users showing greater vulnerability to offloading-related decline (Gerlich, 2025).

Taken together, the evidence supports a qualified version of the cognitive-offloading-cost hypothesis as applied to generative AI. The theoretical expectation that offloading reduces internal encoding of delegated material (Risko & Gilbert, 2016; Sparrow et al., 2011) is borne out in GenAI contexts through weaker delayed retention, reduced neural connectivity during AI-assisted tasks, and lower self-reported critical thinking among frequent users. At the same time, the magnitude and even the direction of these effects depend heavily on how AI is used. This pattern is more consistent with a metacognitive-control account, in which the person's degree of active engagement with AI output-rather than the mere fact of using AI-determines the cognitive consequence, than with a simple technological-determinist account in which any AI use necessarily erodes cognition.

This has two important implications. First, it suggests that the widely used umbrella term “cognitive offloading” may be insufficiently granular for the GenAI era, and that the field is converging on a need to distinguish autonomous from dependent offloading, or offloading from cognitive surrender, as separate constructs with different downstream consequences (Frontiers, 2026). Second, it reframes the practical question away from “should AI be used” and toward “under what conditions of task design, guidance, and user metacognitive engagement does AI support versus supplant learning.” Findings that structured prompting and instructor guidance can substantially attenuate offloading costs are encouraging for educational and clinical training contexts, where GenAI adoption is already outpacing evidence-based guidance.

The literature also reveals a persistent methodological limitation: most studies are cross-sectional, survey-based, or short-duration experiments, and the strongest neurophysiological evidence (Kosmyna et al., 2025) comes from a single, methodologically contested study with a modest sample. Effect sizes for learning outcomes vary widely by discipline, task type, and outcome measure, and few studies have followed users over the months or years needed to assess whether observed short-term effects on retention and connectivity translate into durable changes in skill or expertise. The literature is also skewed toward educational and knowledge-work contexts in higher-income countries, limiting generalizability.

Limitations of the Reviewed Evidence

• Reliance on self-report measures of critical thinking and offloading tendency in several key studies, which are susceptible to social-desirability and introspection biases.
• Small sample sizes in neurophysiological work, limiting statistical power and generalizability.
• Predominantly cross-sectional or short-duration experimental designs, precluding strong causal claims about long-term cognitive change.
• Heterogeneous operationalizations of “offloading,” “dependency,” and “metacognitive engagement” across studies, complicating direct comparison and precluding formal meta-analytic pooling in this review.
• Rapid evolution of GenAI systems themselves, meaning findings based on earlier model generations may not generalize to current or future tools.

The literature on cognitive offloading in the age of generative AI converges on a nuanced but coherent picture. GenAI extends the scope of offloading from discrete, well-defined subtasks to higher-order reasoning, synthesis, and evaluation, and this expansion carries measurable costs for memory retention, self-regulated learning, and markers of executive engagement when AI output is adopted uncritically. These costs are not fixed: they are substantially moderated by instructional structure, task complexity, and-most importantly-by whether the user remains an active, evaluative participant in the AI-assisted process (autonomous offloading) or defers judgment to the system (dependent offloading or cognitive surrender). For educators, clinicians-in-training, and designers of AI-assisted learning tools, the practical implication is that mitigating cognitive cost is less a matter of restricting access to GenAI than of structuring its use to preserve active cognitive engagement-for example, requiring initial unaided effort, scaffolded prompting that demands justification and evaluation, and deliberate practice without AI support at intervals. Future research should prioritize longitudinal designs, objective (non-self-report) measures of executive function and retention, and replication of key neurophysiological findings, in order to determine whether the short-term effects documented to date accumulate into durable changes in memory, learning capacity, and executive function.