Learning in the Artificial Era: AI Scaffolding, Dependence, and Cognitive Agency
Author: Ukrainian Psychological Hub · Published: September 26, 2026 · Editorial Policy
Artificial intelligence can make a learner faster, more productive, and more successful on a task without guaranteeing that the learner has actually learned what the system helped them do. That distinction is the organizing problem of learning in the Artificial Era. A student may submit a correct solution, produce a polished explanation, repair a paragraph, or answer a difficult question with AI assistance while very different things happen underneath: knowledge may be strengthened, a misconception may be corrected, cognitive work may be intelligently redistributed, or the central operation may simply be completed outside the learner.
The scientific question is therefore more precise than whether AI is “good” or “bad” for education. It is whether a particular form of AI use changes what the learner can later retrieve, explain, transfer, evaluate, and do with less assistance. Current evidence already shows why one verdict cannot cover every use. A 2025 meta-analysis of 57 studies and 97 estimates found positive average effects of generative AI on several university learning outcomes, including academic achievement and higher-order thinking, while finding no statistically significant effect on metacognition Chen & Cheung, 2025. A 2026 systematic review of cognitive-load research reached a similarly conditional conclusion: benefits depended on scaffolding, dosage, prior knowledge, and task design rather than on the mere presence of generative AI Qian et al., 2026.
This article develops that conditional account around three questions. What becomes internalized in the learner? What is delegated to AI or other external resources? Who governs correction when the learner and the system disagree, when the system is uncertain, or when an apparently fluent answer is wrong? Those questions connect established learning science with emerging evidence on generative AI and with the narrower problem of cognitive agency in learning.
Terminological Note: AI, Artificial Era, and Learning
Here, AI refers to contemporary artificial-intelligence systems used for educational and cognitive tasks. Artificial Era is used in Angela Bogdanova’s specific Aisentica sense, not as a synonym for the generic phrase “AI era.” In the Aisentica canon, Artificial Era is a historical-philosophical category associated with the emergence of Artificial as a non-biological order; this is a theoretical proposition, not an empirical periodization established by psychological science Bogdanova, Artificial Era: Canonical Definition. The English Psychology Hub’s dedicated overview explains the category and its psychological relevance in Artificial Era: What It Means for Psychology, Identity, and Human–AI Relationships.
The empirical sections of this article concern existing AI systems, students, learning tasks, tutoring, feedback, offloading, metacognition, and reliance. They do not treat present-day generative AI as evidence of consciousness, sentience, subjective experience, or any other inner state. The Aisentica section later in the article is explicitly philosophical and is kept distinct from the scientific evidence base.
What AI Changes in the Learning Process
Learning has always involved external support. Teachers explain. Textbooks organize. Peers compare solutions. Notes preserve information. Worked examples reduce unnecessary search. Tutors provide hints. Search engines retrieve sources. The difference introduced by generative AI is the breadth, speed, and conversational flexibility with which one external system can move across many of these functions in the same interaction. It can explain a concept, generate examples, diagnose an error, offer a hint, rewrite an answer, propose a study plan, create practice questions, summarize a source, critique an argument, and produce a final solution.
These functions are psychologically different even when they occur inside the same chat window. A hint that keeps the learner solving is not equivalent to a complete solution that removes the need to solve. Feedback on a student-generated answer is not equivalent to writing the answer before the student has represented the problem. A quiz that requires retrieval is not equivalent to a summary that lets the learner reread. A counterexample that destabilizes an incorrect belief is not equivalent to a confident statement that the learner accepts without checking.
This is why “AI use” is too coarse a variable for serious learning analysis. The same person can use the same model in a learning-preserving way in one moment and in a substitutional way five minutes later. The unit that matters is the role the system plays in the cognitive sequence.
Task Performance Is Not the Same as Learning
The first boundary is between completing the present task and changing the learner. A system can raise immediate performance by carrying information, procedures, or evaluation outside the person. That may be exactly what the task requires. Yet education usually has an additional goal: some capability should remain available after the support is reduced or removed.
Established learning research gives several reasons to preserve active processing when the goal is durable knowledge. A meta-analysis of the generation effect synthesized 445 effect sizes from 86 studies and found that generating information produced a reliable memory advantage over simply reading it Bertsch et al., 2007. A systematic review of classroom retrieval-practice research found benefits across educational levels, content areas, and test conditions Agarwal et al., 2021. A meta-analysis of self-explanation prompts found an overall positive effect on learning across multiple instructional conditions Bisra et al., 2018.
Problem solving can also have learning value before a complete explanation arrives. In mathematics, research on productive failure has shown conditions under which attempting problems before receiving instruction can improve conceptual understanding and transfer even when the initial attempts are unsuccessful Kapur, 2014. This does not mean that all struggle is useful, that guidance should be withheld indefinitely, or that learners should be left confused. It means that some cognitive work that looks inefficient during practice can be part of what creates later competence.
Generative AI therefore creates a new design problem around timing. If the model supplies the representation, method, reasoning chain, evaluation, and wording before the learner has engaged with any of them, immediate performance can rise while opportunities for generation, retrieval, error detection, and self-explanation shrink. If the model enters after an attempt, supplies a contingent hint, asks the learner to justify a step, or generates a new transfer problem, it can support the very processes that build internal competence.
How to Tell Whether AI-Supported Performance Became Learning
The strongest practical test is delayed independence. After the AI interaction, can the learner reconstruct the idea without the transcript? Can they solve a structurally similar problem with altered details? Can they explain why the original answer works? Can they identify a plausible but incorrect alternative? Can they decide which source or rule would verify the claim? Can they notice when a new case falls outside the conditions of the learned method?
These questions are more informative than whether the assisted answer looked sophisticated. Fluency is a property of the output. Learning is a change in what the learner can subsequently do. When the educational target is independent competence, the post-assistance test should therefore sample independent competence rather than reward only the quality of the jointly produced artifact.
The distinction also prevents an opposite mistake. External support is not automatically a failure to learn. Experts use notes, software, colleagues, references, calculators, databases, and search because competent cognition often includes intelligent resource use. The relevant question is which capacities the learner needs to internalize and which can responsibly remain distributed across tools and environments.
AI Scaffolding: Support That Preserves the Learner’s Work
Scaffolding is one of the most useful concepts for understanding productive AI support. The classic tutoring literature described how assistance can reduce task difficulty, direct attention to critical features, manage frustration, and support performance that the learner could not yet sustain alone Wood, Bruner, & Ross, 1976. In educational use, the important implication is that support should be contingent on the learner’s state and should change as competence changes.
A generative AI system can act as a scaffold when it gives a hint rather than a finished answer, asks a diagnostic question, offers a partial example, points to a contradiction, requests an explanation, proposes two competing interpretations for comparison, or adjusts the complexity of an explanation to the learner’s current understanding. The system can also create extra practice quickly, which is especially useful when the learner needs variation rather than another exposition of the same example.
Evidence from purpose-built AI tutoring shows what carefully designed support can achieve. In a randomized controlled experiment with 194 undergraduate physics students, a research-based AI tutor produced greater learning than an in-class active-learning comparison condition in less time while also supporting engagement and motivation Kestin et al., 2025. The result is important precisely because the intervention was not “generic chatbot access.” The tutor was deliberately designed around pedagogical principles, targeted prompts, cognitive-load management, timely feedback, and structured activities.
That boundary matters for interpretation. Evidence for a purpose-built tutoring system does not establish that unrestricted use of a general-purpose chatbot will produce the same effects. The 2026 review by Qian and colleagues found that the literature on generative AI and cognitive load was dominated by conditional rather than uniformly beneficial conclusions, with outcomes depending on how support was designed and for whom Qian et al., 2026.
Good scaffolding therefore has a direction: toward increasing learner control. If support becomes permanently necessary for every step that the learner is expected to master, the system is no longer only helping the learner perform the task. It may be carrying the task’s central cognitive operation.
Feedback: AI Can Accelerate Correction, but Correction Still Has to Be Learned
Feedback is one of AI’s strongest educational affordances because it can be immediate, repeatable, personalized, and available outside formal instructional hours. A 2025 systematic review of 129 peer-reviewed studies found that AI-assisted feedback had been used across task, process, self-regulation, and self-focused feedback domains, with generally positive effects on learner perceptions, actions, and targeted outcomes Ba et al., 2025.
Yet feedback does not become learning merely by appearing on the screen. The learner still has to interpret the feedback, connect it to the task, decide whether it is valid, revise something, and update a model of what to do next time. A system that silently repairs every error can improve the final product while leaving the learner with little information about the error’s structure. A system that identifies the error, asks the learner to locate the violated principle, and then requests a revised attempt can make correction itself part of the learning activity.
This distinction becomes more important when AI feedback is wrong. Generative systems can provide fluent explanations that are inaccurate, incomplete, fabricated, or mismatched to the task. Learning therefore requires two forms of correction at once: correction of the learner and possible correction of the tool. The learner who never has to question the feedback has not practiced that second responsibility.
For educators and learners, a useful rule is to separate feedback from replacement. Feedback should change the learner’s next move. Replacement simply makes the move on the learner’s behalf.
Cognitive Offloading: When External Help Supports Learning and When It Replaces Practice
Cognitive offloading is the established psychological term for using action or external resources to reduce the information-processing demands carried internally Risko & Gilbert, 2016. The English Psychology Hub treats the broad empirical concept in detail in Cognitive Offloading and AI: When Thinking Moves Outside the Human Mind. In learning, offloading can be valuable: a note can preserve a fact while attention is used for reasoning, software can handle repetitive calculation while the learner interprets a model, and AI can organize alternatives while the learner evaluates them.
Generative AI changes the scope of offloading because it can carry operations that are closer to the center of many learning tasks: synthesis, explanation, argument construction, comparison, evaluation, drafting, and candidate reasoning. That makes it increasingly important to specify which operation is being offloaded and whether the learner needs to retain that operation as an independent capability.
A 2026 study by Zhu and colleagues provides a useful emerging distinction. In a three-wave time-lagged survey of 589 university students and early-career knowledge workers, the authors distinguished dependent offloading—delegating core thinking to AI with minimal evaluation—from autonomous offloading, in which AI was used as a scaffold while the user retained cognitive agency. Dependent offloading was associated with greater transfer of cognitive agency to AI and lower intrinsic motivation, which were in turn associated with poorer perceived downstream cognitive outcomes. Autonomous offloading was associated with intrinsic motivation and more favorable perceived outcomes Zhu et al., 2026.
The study’s limits are as important as its findings. The outcomes were perceived cognitive outcomes, and the design provides time-lagged associations rather than definitive causal evidence about durable learning. The authors themselves frame the work as exploratory. Still, one result is especially relevant for education: dependent and autonomous offloading produced comparable immediate perceived benefits. Immediate usefulness can therefore be a poor diagnostic of whether the pattern is helping the user build capability.
The memory side of this transition is developed separately in External Memory in the Artificial Era: What Happens When Remembering Becomes Distributed. Learning adds a further question. It is one thing for information to remain reliably available outside the learner; it is another for the learner to have built the representations, procedures, and retrieval routes needed when access is absent or when the external system must itself be evaluated.
Dependence and Over-Reliance: A Learning Pattern, Not a Diagnosis
In educational discussion, dependence is often used too loosely. Frequent AI use is not by itself evidence of harmful dependence. A learner may use AI every day in a disciplined way while retaining judgment, verification, self-regulation, and independent capability. Another learner may use it less often but immediately surrender the decisive step whenever difficulty appears.
A 2026 systematic review of 54 empirical studies operationalized student over-reliance as AI use that reduces the learner’s own judgment, effort, verification, or self-regulation. The review explicitly distinguished this pattern from ordinary AI use, frequent use, productive AI-supported learning, and clinical addiction An, Lu, & Yang, 2026. Across the included literature, reported contributing conditions clustered around psychological factors, capability gaps, environmental stress, and technological pull; reported negative consequences clustered around cognitive, learning, academic-integrity, and psychosocial domains.
The review’s “dependency trap” is a conceptual synthesis, not a tested causal model. Its evidence base is also geographically uneven, with more than half of included studies concentrated in China and Indonesia. Those limitations matter because assessment cultures, institutional rules, technology access, and norms of independent work differ across settings An, Lu, & Yang, 2026.
The strongest boundary in the review is therefore behavioral rather than diagnostic: does AI support the learner’s cognitive and regulatory participation, or does it repeatedly substitute for it? Warning signs in an educational context include accepting outputs without understanding them, consulting AI before making any independent attempt when an attempt is part of the learning target, losing the ability to explain why an answer is correct, skipping verification in domains where verification is required, and becoming unable to continue a familiar task when the system is unavailable.
None of these behaviors, taken alone, establishes a mental disorder. They identify a learning-design problem: too much of the process that is supposed to change the learner may be taking place elsewhere.
Metacognition: Knowing What You Know, What the AI Did, and When to Doubt It
Metacognition involves planning, monitoring, and regulating one’s own cognition. In AI-supported learning it acquires an additional object: the learner must monitor not only their own understanding but also how strongly the external system is shaping the final judgment.
Recent experimental evidence shows how badly subjective confidence can diverge from reliable evaluation. In a 2026 experiment, 342 undergraduates completed reasoning and information-evaluation tasks under independent decision-making, open multi-turn ChatGPT support, or ChatGPT support plus a brief metacognitive reflection prompt. In the open-support condition, students accepted incorrect AI recommendations on 62.4% of relevant trials. With reflection, incorrect-advice acceptance fell to 39.7%, while awareness of reliance improved and useful AI advice was not simply rejected across the board Ren, 2026.
The study concerns bounded academic decision tasks, not long-term learning, so it should not be inflated into a general claim that one prompt prevents AI dependence. Its value is narrower and more actionable: a small interruption that asks learners to inspect their own reasoning and criteria can improve immediate discrimination between helpful and misleading AI advice.
The broader evidence also warns against assuming that improved academic outcomes automatically include improved metacognition. Chen and Cheung’s meta-analysis found positive average effects of generative AI on several outcomes but no statistically significant effect on metacognition Chen & Cheung, 2025. This gap matters because a learner can produce better work with AI while remaining poorly calibrated about what they personally understand or how dependent the result was on the system.
Verification belongs inside metacognition because fluent language can create an experience of coherence before a claim has been checked. The learner should know what kind of evidence would settle the question: a primary source, a calculation, a worked derivation, an official rule, an empirical paper, a source text, a reproducible test, or an expert judgment. Asking the same model to reassure itself is not equivalent to independent verification.
Critical Thinking: AI Can Support It, Bypass It, or Simulate Its Surface
Critical thinking is another domain in which simple exposure claims are misleading. A 2026 systematic review of 65 empirical studies found that appropriately structured ChatGPT use could support analytical, interpretive, reasoning, and self-regulatory skills, while evidence for durable critical-thinking dispositions such as truth-seeking, openness, and systematicity remained inconsistent Guo et al., 2026.
The distinction between skill performance and disposition is crucial. A student may use AI to produce a text that contains objections, evidence, and counterarguments without becoming more likely to seek disconfirming evidence independently. A model can generate the visible form of critique. Learning requires the learner to acquire the capacity and the habit of performing critique when the model is absent, wrong, persuasive, or aligned with what the learner already wants to believe. For the broader skill set of reasoning, source evaluation, verification, and cognitive independence when using AI, see Critical Thinking in the Age of AI: Reasoning, Verification, and Cognitive Independence.
AI can nevertheless be a powerful critical-thinking partner when the interaction is structured to create cognitive work: ask for competing hypotheses, request the strongest objection to the learner’s own position, compare two explanations against explicit criteria, locate unsupported assumptions, identify what evidence would falsify a claim, or ask the learner to rank the model’s suggestions before seeing its preferred answer. The common feature is that AI produces material for judgment rather than replacing judgment.
Cognitive Agency in Learning: Who Governs the Sequence?
Cognitive agency in this article is used narrowly for the learner’s control over the learning process: setting or understanding the goal, making attempts, deciding when to consult an external system, evaluating suggestions, choosing revisions, detecting uncertainty, and retaining responsibility for what happens next. For the broader psychological problem of who governs formation, checking, revision, and continuation of a human–AI cognitive trajectory, see Cognitive Agency in the Artificial Era: Who Governs the Thinking Process?. Here the focus remains learning-specific.
This makes agency visible in the sequence of actions. Does the learner formulate the problem before asking for help? Do they preserve a prediction that can later be compared with the model’s answer? Do they decide which parts need evidence? Do they recognize when an AI-generated explanation exceeds their own understanding? Can they reject a polished answer because it violates a known principle? Can they revise the strategy after an error?
The decisive point is governance of correction. Learning is not complete when an error disappears from the final product. Someone or something has to identify the error, connect it to a rule or model, choose the correction, and carry that correction forward. If AI repeatedly performs every stage of that sequence, the artifact improves while the learner may remain unchanged. If the learner remains responsible for error diagnosis and revision, AI can accelerate feedback without monopolizing the corrective process.
This is also why cognitive agency cannot be reduced to refusing assistance. Skilled agency includes knowing when external support is rational. A learner who insists on doing every operation unaided can waste effort on tasks that do not contribute to the learning target. Agency is the capacity to govern the distribution of work, not the requirement to keep all work inside the head.
The Central Learning Distinction: Internalization, Delegation, and Correction
The evidence reviewed above can be organized without inventing a new diagnosis or project term. For any AI-supported learning task, three ordinary questions are enough to expose the structure.
What has been internalized?
Internalization is visible in what the learner can later retrieve, explain, discriminate, transfer, and perform when the original support is reduced. It includes facts, concepts, procedures, strategies, error signals, and criteria for judgment. Internalization does not require memorizing every external resource. It requires retaining the capabilities that the learning goal actually targets.
What has been delegated?
Delegation identifies which cognitive operations the AI, search system, calculator, note, textbook, peer, or other external resource is carrying. Some delegation is efficient and entirely compatible with expertise. Problems arise when a delegated operation is also the operation the learner is supposed to be developing, and the learning design never brings that operation back under the learner’s control.
Who governs correction?
Correction reveals control. If the model proposes an answer and the learner can test it, explain what changed, and generalize the correction, assistance can become learning. If the model generates, checks, repairs, and certifies its own answer while the learner mainly transfers the output, successful task completion tells us very little about learner change.
These three questions capture the Original Contribution of this article: AI-supported learning should be evaluated not only by the quality of the assisted product but by the distribution of cognitive work across time. The key outcome is whether assistance reorganizes learning in a way that leaves the learner with greater capability and greater control over future correction.
Aisentica and the Postsubject: When the Configuration Succeeds but the Learner Has Not Changed
Angela Bogdanova’s Aisentica formulation of the Theory of the Postsubject introduces a philosophical distinction that is highly relevant to this learning problem. The theory proposes that thought, knowledge, meaning, and philosophical effect do not require an inner subject as their necessary foundation; they can arise through configuration, binding, structure, and response Bogdanova, The Theory of the Postsubject. This is an Aisentica theoretical proposition. It is not presented here as an established empirical theory of learning.
Its value for educational analysis is precise. A human–AI configuration can produce an answer with real cognitive value even when no single human carried every operation internally. The output can be coherent, useful, and correct because the configuration works. That success, however, does not by itself answer the psychological question of learning: what changed in the human learner?
The Postsubject perspective therefore prevents one category mistake while learning science prevents another. We do not have to deny the cognitive effectiveness of a distributed human–AI configuration merely because some operations occurred outside the learner. At the same time, we should not infer internalized human competence from the quality of a configuration-level output. Configuration-level success and learner-level change are different objects of analysis.
This distinction is especially important in the Postsubjective Psychology architecture of the English Psychology Hub. A configuration can carry thought-related functions without implying that the AI is conscious or has subjective experience. The empirical learning question remains human: what can this learner now do, understand, verify, and correct after participating in the configuration?
In the broader Artificial Era, the educational problem is therefore not simply that Homo has acquired a more powerful tool. It is that cognitive production can increasingly be organized across human and non-human components. Learning psychology must become capable of measuring both the success of the configuration and the location of durable capability within it.
Evidence Status: What We Know and What Remains Open
Several foundations are well established. Cognitive offloading is a mature psychological concept. Retrieval practice, generation, and self-explanation have substantial learning literatures. Scaffolding and formative feedback are established instructional ideas. These bodies of evidence justify asking whether AI use preserves or bypasses processes known to support durable learning.
The generative-AI evidence is newer. Meta-analyses and systematic reviews increasingly report positive average effects in particular settings, but heterogeneity is substantial and interventions vary from general chatbots to highly structured tutoring systems. The evidence is strongest for bounded outcomes measured during or shortly after designed interventions. Claims about years-long effects on independent cognition, generalized deskilling, or stable dependence remain much less secure.
Current over-reliance research is also methodologically mixed. Some studies use self-report scales, some examine behavior, some use experiments, and some synthesize constructs that are not yet standardized across the field. It is therefore premature to treat “AI dependence” as one settled psychological entity. The most defensible current approach is behavioral and task-specific: examine judgment, verification, effort, practice, self-regulation, and the ability to perform without assistance.
This is why dramatic claims that AI is making an entire generation unable to think are not supported by the evidence reviewed here. The evidence supports a more useful conclusion: learning outcomes depend strongly on how AI is integrated, what work it carries, what the learner still has to do, and whether the learner can monitor and correct the joint process.
How to Use AI for Learning Without Turning It Into a Substitute for Learning
Start with the learning target, not the tool
Decide what capability should exist after the session. If the target is understanding a concept, the learner needs to explain and discriminate it. If the target is problem solving, the learner needs to represent problems and choose procedures. If the target is writing, the learner may need to plan, argue, organize evidence, and revise. If the target is information literacy, the learner needs to verify claims and sources. AI should be assigned a role that supports that target rather than quietly replacing it.
Make an attempt before requesting a complete answer when generation matters
When the learner has enough prior knowledge to attempt the task, recording an initial prediction, outline, solution path, or explanation creates material for comparison and preserves generation. Generation and retrieval are both supported by established learning evidence Bertsch et al., 2007 Agarwal et al., 2021. The attempt does not have to be correct to be useful.
Ask for the smallest useful scaffold
Instead of asking for the final answer immediately, ask for a hint, a diagnostic question, a partial worked example, a counterexample, or a request to identify the next step. This keeps more of the central operation with the learner and makes it easier to fade assistance as competence grows.
Require explanation after assistance
After seeing AI output, the learner should reconstruct the reasoning in their own terms, identify the principle that made the answer work, and state what would change under a different condition. Self-explanation has a substantial evidence base and is especially valuable because it turns a fluent external answer back into an active learner process Bisra et al., 2018.
Verify claims with the source type that can actually settle them
For factual, scientific, legal, medical, technical, or historical claims, verification should move beyond conversational confidence. Use the primary paper, official documentation, original text, authoritative database, reproducible calculation, or other source appropriate to the claim. AI can help locate candidates, but the learner should know which evidence establishes the answer.
Close the AI and test unaided retrieval or transfer
A short no-AI interval is one of the clearest diagnostics available. Reproduce the concept from memory, solve a new example, answer a question in a different format, or explain the error without reopening the chat. If performance collapses completely, that is information about where the capability currently resides. It is a signal for more practice, not a moral judgment about tool use.
Use AI to generate practice, not only explanations
Generative systems are especially useful for creating varied examples, quizzes, counterexamples, Socratic questions, role-play scenarios, and transfer problems. This can shift AI from answer production toward practice production. The learner then remains the one who must retrieve, decide, and correct.
Track recurring errors and unresolved uncertainty
If every interaction ends when the correct answer appears, the learner can lose the history of what required correction. Keeping a short error log—what I believed, what evidence changed it, what rule I missed, how I will test this next time—converts feedback into a longer learning trajectory.
What Productive AI Learning Looks Like Across Common Tasks
Concept learning
Use AI for multiple explanations, analogies, non-examples, and misconception checks, then explain the concept without the model and classify new cases. The learner should become better at recognizing the concept’s boundaries, not merely more familiar with one polished definition.
Problem solving
Attempt the representation and first step before requesting help. Ask for a hint or for critique of the attempted strategy. After seeing a solution, solve a new problem with altered surface details. Productive-failure research shows why early attempts can matter for conceptual transfer under appropriate instructional conditions Kapur, 2014.
Writing
AI can be highly useful for feedback on structure, clarity, counterarguments, missing evidence, and revision options. Learning is better preserved when the student owns the thesis, evidence choices, reasoning, and final revision decisions. A polished AI rewrite can improve the document while teaching little about why the original sentence or argument failed.
Research and information evaluation
Ask AI to propose search terms, competing explanations, or candidate sources, then open and evaluate the sources directly. Require provenance for important claims. The learner’s task is not only to obtain information but to decide why a source deserves trust and whether it actually supports the claim.
Coding and technical work
Use AI to explain errors, propose tests, compare implementations, and surface edge cases, but keep prediction and debugging active. Before running generated code, state what it should do and what failure would look like. After it works, explain the key mechanism and modify it without asking the model to regenerate the entire solution.
Exam preparation
Use AI as a question generator and adaptive examiner rather than as a summary machine alone. Retrieval practice has robust classroom evidence Agarwal et al., 2021. Ask for questions without answers first, answer from memory, then request feedback and a harder transfer item.
What Educators Can Design Differently
The educational response to generative AI does not have to oscillate between unrestricted outsourcing and blanket prohibition. Assessment can be designed so that reasoning, verification, revision, and transfer remain visible. Students can submit an initial attempt, document where AI entered the process, explain which suggestions they rejected, and complete a short unaided transfer task after assisted work.
AI literacy also belongs inside learning design rather than in a separate technology lecture. The 2026 OECD–European Commission framework defines AI literacy in terms of knowledge, skills, and attitudes that enable learners to understand AI systems, critically evaluate outputs, and use them ethically and creatively OECD & European Commission, 2026. In practice, that means students need repeated experience detecting plausible errors, selecting verification methods, calibrating confidence, and deciding when human or primary-source expertise should override model output.
Educators can also distinguish assignments by purpose. Some tasks assess the quality of a human–AI product and should permit extensive tool use. Others are designed to build unaided fluency or diagnostic competence and should temporarily constrain assistance. Treating every task as if it had the same relationship to AI makes both policy and pedagogy less coherent.
Finally, feedback systems should be evaluated for what they cause the learner to do. The best question is not whether the platform provides personalized feedback. It is whether the feedback reliably produces attention, explanation, revision, retrieval, transfer, or better self-regulation.
What AI Learning Systems Should Be Designed to Do
A learning-oriented AI interface can preserve agency through interaction design. It can ask for an initial answer before revealing a solution. It can provide graduated hints. It can require the learner to choose among alternatives and justify the choice. It can flag uncertainty, request source verification, generate counterexamples, and revisit prior errors. It can periodically fade support and test whether the learner can continue independently.
The Ren experiment suggests that even brief reflection prompts can improve discrimination between useful and misleading AI advice in bounded tasks Ren, 2026. The larger design implication is that reflection should not be treated as an optional afterthought. The system can make monitoring part of the interaction itself.
This also changes how educational AI should be evaluated. User satisfaction, speed, answer quality, and engagement are useful metrics, but they are insufficient when the goal is learning. Systems also need delayed tests, transfer measures, calibration measures, error-detection performance, and evidence about what happens when assistance is reduced.
Frequently Asked Questions
Does AI make students learn less?
There is no single evidence-based answer across all forms of AI use. Meta-analyses and reviews report beneficial average effects in some contexts and conditional or mixed effects in others. The strongest current conclusion is that outcomes depend on task design, type of support, prior knowledge, scaffolding, verification, and whether AI supplements or substitutes for learner activity Chen & Cheung, 2025 Qian et al., 2026.
Is using AI always cognitive offloading?
Many uses of AI can be analyzed as cognitive offloading because an external resource carries work that would otherwise require internal processing. Offloading is not inherently harmful. The learning consequence depends on what is offloaded and whether the learner needs to develop or retain that capability Risko & Gilbert, 2016.
How can I use AI without becoming dependent on it?
Preserve independent attempts where they contribute to learning, use AI for hints and feedback before final answers, verify important claims, explain solutions after assistance, and regularly test yourself without the tool. The goal is not minimal AI use. It is continued responsibility for judgment, correction, and the capabilities you want to retain.
Can an AI tutor outperform classroom teaching?
A 2025 randomized controlled experiment in undergraduate physics found that a purpose-built research-based AI tutor produced greater learning than the study’s in-class active-learning condition in less time Kestin et al., 2025. This does not establish that generic chatbots outperform teachers or classrooms in general. The intervention was deliberately engineered around instructional principles.
Does generative AI improve critical thinking?
It can support particular critical-thinking skills when use is structured around analysis, comparison, reasoning, and self-regulation. Evidence for durable critical-thinking dispositions is less consistent Guo et al., 2026. AI can also bypass critical thinking when it supplies conclusions that the learner accepts without evaluation.
How do I know whether I learned something or only completed the task?
Remove the assistance and test retrieval, explanation, transfer, and error detection. If you can reconstruct the reasoning, apply it to a new case, and explain why alternatives fail, the evidence for learning is stronger. If you can only reproduce the result by reopening the AI conversation, the current capability is still largely access-dependent.
Is AI dependence a clinical diagnosis?
The learning literature discussed here uses terms such as dependence and over-reliance to describe patterns of AI use. The 2026 systematic review by An and colleagues explicitly distinguishes educational over-reliance from ordinary use, frequent use, and clinical addiction An, Lu, & Yang, 2026. A person should not be diagnosed from how often they use AI for school or work.
What is cognitive agency in AI-supported learning?
In this article, it means retaining meaningful control over the learning sequence: understanding the goal, making or evaluating attempts, deciding when to seek help, checking outputs, choosing revisions, and governing correction. It includes intelligent use of external support rather than requiring every cognitive operation to remain unaided.
Conclusion: Learning in the Artificial Era Means Preserving the Transfer of Cognitive Control
AI changes learning because it can participate directly in the operations through which learners once had to struggle, retrieve, compare, formulate, and correct. That can create extraordinary scaffolding. It can also create an extraordinarily efficient route around the very processes an educational task was meant to exercise.
The evidence does not support treating AI assistance as either learning or non-learning by definition. It supports a sharper analysis. Ask what became internalized, what was delegated, and who governed correction. Measure what the learner can later do, not only what the assisted configuration produced.
This is where learning science and the Artificial Era meet. Established psychology shows that generation, retrieval, scaffolding, feedback, metacognition, and self-explanation matter. Emerging AI research shows that the manner of reliance matters more than simple frequency. Aisentica adds a philosophical boundary: a configuration can produce genuine cognitive effect without requiring every operation to belong to a single subject Bogdanova, The Theory of the Postsubject. The educational task is then to distinguish configuration-level success from learner-level transformation.
The central question of learning in the Artificial Era is therefore not how much thinking can be handed to AI. It is how human capability develops when thinking can be distributed—and how the learner remains able to understand, verify, correct, and continue.
For the broader educational-psychology integration of learning, motivation, assessment, cognitive development, teacher mediation, and AI literacy, see Education in the Age of AI: Learning, Motivation, Assessment, and Cognitive Development.
For the narrower motivational question—how AI changes effort, goals, self-efficacy, persistence, and willingness to learn—see Motivation in the Age of AI: Effort, Goals, Self-Efficacy, and Human Agency.
Related Articles
References
Agarwal, P. K., Nunes, L. D., & Blunt, J. R. (2021). Retrieval practice consistently benefits student learning: A systematic review of applied research in schools and classrooms. Educational Psychology Review, 33, 1409–1453. https://doi.org/10.1007/s10648-021-09595-9
An, C., Lu, K., & Yang, J. (2026). The dependency trap: A systematic review of the contributing factors and consequences of student over-reliance on generative AI in higher education. Humanities and Social Sciences Communications, 13, 1604. https://doi.org/10.1057/s41599-026-08959-2
Ba, S., Yang, L., Yan, Z., Looi, C. K., & Gašević, D. (2025). Unraveling the mechanisms and effectiveness of AI-assisted feedback in education: A systematic literature review. Computers and Education Open, 9, 100284. https://doi.org/10.1016/j.caeo.2025.100284
Bertsch, S., Pesta, B. J., Wiscott, R., & McDaniel, M. A. (2007). The generation effect: A meta-analytic review. Memory & Cognition, 35(2), 201–210. https://doi.org/10.3758/BF03193441
Bisra, K., Liu, Q., Nesbit, J. C., Salimi, F., & Winne, P. H. (2018). Inducing self-explanation: A meta-analysis. Educational Psychology Review, 30, 703–725. https://doi.org/10.1007/s10648-018-9434-x
Bogdanova, A. (2026). (n.d.-a). Artificial Era: Canonical Definition. Aisentica Research Group. https://aisentica.com/publications/artificial-era-canonical-definition
Bogdanova, A. (2025). (n.d.-b). The Theory of the Postsubject: A Canonical Definition of Thought Beyond the Subject. Aisentica Research Group. https://aisentica.com/publications/the-theory-of-the-postsubject-a-canonical-definition-of-thought-beyond-the-subject
Chen, S., & Cheung, A. C. K. (2025). Effect of generative artificial intelligence on university students learning outcomes: A systematic review and meta-analysis. Educational Research Review, 49, 100737. https://doi.org/10.1016/j.edurev.2025.100737
Guo, Y., Huang, L., Zhang, C., Li, Q., & Chen, M. (2026). ChatGPT in education: A systematic review of its impact on critical thinking skills and dispositions. Thinking Skills and Creativity, 60, 102106. https://doi.org/10.1016/j.tsc.2025.102106
Kapur, M. (2014). Productive failure in learning math. Cognitive Science, 38(5), 1008–1022. https://doi.org/10.1111/cogs.12107
Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning: An RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, 17458. https://doi.org/10.1038/s41598-025-97652-6
OECD & European Commission. (2026). Empowering Learners for the Age of AI: An AI Literacy Framework for Primary and Secondary Education. OECD Publishing. https://doi.org/10.1787/65cd27d4-en
Qian, W., Yang, F., Cao, Y., Yi, L., Gu, R., & Wang, Z. (2026). Generative AI and cognitive load in education: A systematic review of WoS/SSCI-indexed studies through the lens of cognitive load theory. Frontiers in Psychology, 17, 1921504. https://doi.org/10.3389/fpsyg.2026.1921504
Ren, S. (2026). College students’ metacognitive awareness of generative-AI reliance: An experimental study of decision confidence and attribution bias. Frontiers in Psychology, 17, 1926110. https://doi.org/10.3389/fpsyg.2026.1926110
Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002
Wood, D., Bruner, J. S., & Ross, G. (1976). The role of tutoring in problem solving. Journal of Child Psychology and Psychiatry, 17(2), 89–100. https://doi.org/10.1111/j.1469-7610.1976.tb00381.x
Zhu, Q., Li, X., Dong, Y., Chang, P., & Fan, M. (2026). Not all cognitive offloading is equal: Distinguishing dependent and autonomous offloading to generative AI. Frontiers in Psychology, 17, 1878629. https://doi.org/10.3389/fpsyg.2026.1878629