The strongest individual studies point in opposite directions. In a field experiment with nearly a thousand secondary mathematics students, access to a standard chatbot during practice dramatically improved performance on the practice problems themselves, yet lowered performance on a subsequent exam taken without AI relative to peers who never had access. Students treated the tool as a crutch, reproducing its answers, errors included, without engaging the underlying mathematics, and emerged believing they had learned as much as their classmates even as they performed worse. Meanwhile, students who researched a socio-scientific issue with a large language model experienced significantly lower cognitive load than peers using a conventional search engine but produced lower-quality reasoning in their conclusions. Yet in a randomized trial in undergraduate physics, an AI tutor deliberately engineered around the same pedagogical principles as comparison lessons produced greater gains on post-lesson tests, in less time and with higher engagement, than a well-implemented active-learning class. The same technology, studied with comparable rigor, harms learning in one configuration and outperforms exemplary instruction in another.

A new theoretical review published in Educational Psychology Review argues that this contradiction is not noise to be averaged away but the expected signature of a process whose outcome depends on how cognitive labor is divided between learner and machine. The author, Daiki Nakamura of the University of Miyazaki, contends that the question of whether generative AI helps or harms learning is under-specified as posed, because it treats a relational, design-dependent outcome as a fixed property of the tool. His augmentation–erosion framework, integrating research on cognitive offloading, distributed cognition, cognitive load theory, and self-regulated learning, holds that the educational effect of AI is largely a property of the partition of cognitive labor: which cognitive work is done by the student and which is handed off, under what conditions, and with what regulation.

The central principle is elegant. Offloading a cognitive process to AI is augmentative when the process handed off is peripheral to the current learning target: discharging it removes load that would otherwise compete for limited working-memory resources, freeing them for the schema-constructing processing in which learning consists. Offloading is erosive when the process handed off is itself constitutive of the learning target, bypassing the very processing that the learning was supposed to be. Put simply, augmentation frees the learner to do the work that matters; erosion does the work that mattered in the learner’s place. Crucially, erosion can take two forms: acquisition erosion, the failure to build a competence one does not yet have, and maintenance erosion, the decay of a competence already acquired when prolonged offloading removes the exercise on which retention depends.

The framework supplies criteria for locating operations along a graded peripheral–constitutive continuum before instruction begins. An operation is constitutive to the extent that it, or the reasoning it instantiates, is named in the learning goal, and to the extent that a criterion assessment—an unaided task at a delay—would require the learner to perform it. A practical test the learner can ask is: would I have to do this myself on the transfer task? Consider a tenth-grade class modeling bacterial growth. Entering data and plotting may be handed off with benefit; selecting and justifying a model family, and interpreting fitted parameters, must be retained. What makes STEM a strategically revealing case is a structural mismatch: with a calculator, the temptation is to offload arithmetic, which is peripheral to most modern mathematics education. With generative AI, the temptation is to offload the reasoning, the modeling, the argument, the proof—the targets themselves.

The unit of analysis is the episode, not the act. Three students can issue identical requests and diverge completely. One hands the machine data entry, plotting, and parameter fitting, then selects the model family and interprets the growth rate herself—an augmentative partition. Another pastes the data, asks for the best model with a justification, and transcribes the answer—the crutch dynamic documented in the mathematics experiment, at high risk of acquisition erosion. A third issues the same request but then treats the output as a worked example: checking the exponential fit against residuals, explaining why a linear model fails, and later re-deriving the doubling time unaided. The initiating act is identical; what differs is the within-episode processing that follows, and with it the predicted outcome. The same request can open an erosive or an augmentative episode.

Perhaps the framework’s most unsettling contribution is its account of the metacognitive trap. Regulation depends on metacognitive monitoring whose signal is the learner’s sense of their own understanding, and the fluency of generative AI can corrupt that signal by manufacturing an ease that is readily mistaken for comprehension. Judgments of learning follow surface cues such as fluency and ease of processing rather than direct readouts of memory strength. Fluent output inflates those judgments; inflated judgments bias control decisions toward accepting the output and offloading the next step as well; biased control reduces engagement with the focal processes, lowering unaided learning even as assisted performance rises. Empirical signatures are accumulating: AI support has been shown to raise task performance while degrading the accuracy of performance self-assessment, and to improve immediate products without improving knowledge gain while reshaping self-regulatory processes—a pattern its researchers call metacognitive laziness. The consequence is that erosion, when it occurs, tends to be invisible to the person it is happening to, and can become self-reinforcing: apparent success encourages further reliance, which deepens the erosion, which continues to feel like success.

Against this, the framework proposes offloading regulation as a learning-oriented dimension of AI literacy: the metacognitive and epistemic competence to govern the partition of cognitive labor in the service of learning goals. It comprises four coordinated capacities. Partition discernment is the ability to distinguish operations that may be offloaded with benefit from those that must be retained. Calibrated monitoring is the capacity to track actual understanding against the illusion of understanding that AI fluency manufactures. Disciplinary verification evaluates AI outputs against the epistemic standards of the field—whether a mechanism is chemically plausible, a proof step valid, a model’s assumptions defensible—rather than against surface coherence. Re-internalization converts offloaded work back into one’s own competence through re-derivation, self-explanation, or unaided reattempt. Notably, higher self-assessed AI literacy has been associated with worse metacognitive accuracy, underscoring that this capacity is specific and trainable rather than a by-product of exposure.

Design is the system-side lever, and the evidence for it is striking. In the mathematics experiment, a tutor configured to provide teacher-crafted hints rather than solutions kept the target process with the student and eliminated the learning penalty that the answer-giving interface produced. Sequencing matters too: requiring learners to generate their own attempt before consulting AI produced no advantage on the assisted task itself but superior performance on a subsequent unaided task. Assessment operates through incentives: where assessment rewards the finished product, offloading the target process becomes rational; where it demands evidence of unaided capability, it realigns what it is rational to offload. On the contested question of whether novices should use generative AI at all, the framework agrees that prior knowledge is foundational to verification but prescribes external regulation of the partition rather than exclusion—tools configured to offer hints and worked examples rather than finished solutions, with protection progressively handed back as competence grows.

The equity implications invert a common expectation. If access to AI were what determined learning, widening access would tend to narrow gaps. But the framework locates a key determinant in the competence to regulate the partition, which depends on prior disciplinary knowledge and metacognitive skill, both unequally distributed. Learners with stronger preparation are positioned to offload peripheral work, verify what the machine returns, and accelerate; learners with weaker preparation are positioned to offload the target process, accept what they cannot evaluate, and stall—all under identical access to the same tool. Access is necessary, because the competence cannot be practiced without it, but insufficient, because unregulated access stratifies by the preparation learners bring. The author offers this as an explicit, testable hypothesis rather than an established finding, and frames the cultivation of offloading regulation as a matter of equity as well as effectiveness. The promise of these tools, the argument concludes, will not be realized by deciding how much to use them. It will be realized, if at all, by learning—and by teaching—how to divide the thinking.

Subject of Research: Cognitive offloading and its regulation in generative AI-supported STEM learning

Article Title: Augmentation or Erosion? Regulating Cognitive Offloading in Generative AI-Supported STEM Learning

Article References: Nakamura, D. (2026). Augmentation or Erosion? Regulating Cognitive Offloading in Generative AI-Supported STEM Learning. Educational Psychology Review, 38(1), Article 124. https://doi.org/10.1007/s10648-026-10224-6

Image Credits: AI Generated

DOI: 10.1007/s10648-026-10224-6

Keywords: generative AI, cognitive offloading, STEM education, cognitive load theory, self-regulated learning, metacognition, AI literacy, augmentation, erosion, educational psychology, learning outcomes, instructional design