What the Shortcut Changes: Cognitive Effort, Memory, and the Learner’s Agency
Generative AI changes learning not simply by adding an answer source, but by changing who performs the mental steps between a question and an understanding. This chapter focuses on cognitive offloading—using an external aid to reduce mental work—and on the point where useful assistance becomes substitution. The aim is practical: to help the reader inspect an AI interaction and see which actions it preserves, which it removes, and what that means for independent judgment.
The distinction matters because “less effort” can mean two different things. Removing needless difficulty can make room for understanding. Removing the effort that creates understanding can make a task feel easy while leaving the learner less prepared for the next one.
When convenience removes the practice
Cognitive offloading is not automatically a problem. People use notes, calculators, diagrams, search engines, and conversations with other people to hold information outside their heads or reduce the burden of a task. A map can spare someone from memorising every street while helping them reach a destination. In learning, an external aid can free attention from a minor obstacle so that a student can focus on a more important relationship.
The educational risk begins when the aid does not merely reduce an incidental burden. It takes over the very activity the learner needs to practise. If the goal is to learn how to construct an argument, having AI correct a spelling error may be peripheral assistance. Having AI choose the claim, select the evidence, connect the reasoning, and write the conclusion is closer to replacing the learning task. The final text may improve, but the student has had fewer opportunities to build the judgment that the assignment was supposed to exercise.
A useful way to see the difference is to follow the hidden steps inside an apparently simple answer. A learner facing a difficult question normally has to retrieve relevant knowledge, identify what they do and do not understand, form a tentative explanation, compare it with evidence, notice errors, and revise. Those steps are not always comfortable. They can involve uncertainty, failed attempts, and the slow discovery that an attractive answer does not actually explain the problem. That friction is often part of the learning process.
A general-purpose chatbot can interrupt the sequence at several points. It can diagnose the problem before the learner has tried to diagnose it, supply an explanation before the learner has formed a hypothesis, synthesise sources before the learner has compared them, or produce a finished response before the learner has organised an argument. The student may still read the result, but reading a solution is not equivalent to generating and testing one. The missing action is not merely time spent typing. It is the mental work that links information to a usable structure.
This is why cognitive load must be handled carefully. Some load is wasteful: confusing instructions, an inaccessible interface, or a calculation that distracts from the concept can consume attention without teaching anything. A well-designed tool can reduce that burden. But the effort involved in retrieving an idea, explaining why it follows, or deciding whether evidence supports a claim is not necessarily waste. If AI removes that effort too early, it may reduce the learner’s opportunity to encode, organise, and test the knowledge.
The OECD Digital Education Outlook 2026 describes this as the difference between a scaffold and a shortcut. A scaffold gives temporary support that helps a learner perform a difficult action and can be withdrawn as competence grows. A shortcut reaches the visible answer while bypassing the action. The two may look similar from the outside: both can produce a correct response. Their consequences diverge when the student must work without assistance.
Consider two ways of using AI to study a concept. In the first, a student asks, “Explain this chapter in simple terms,” reads the response, and moves on. The interaction may be useful for orientation, but it gives the student little reason to retrieve the central idea, distinguish it from a neighbouring idea, or test whether the explanation fits a new case. In the second, the student first writes a short explanation, asks the system to identify gaps without rewriting it, checks those suggestions against the source, and then revises the explanation in their own words. The second exchange is not guaranteed to teach more, but it keeps diagnosis, comparison, and ownership visible. It gives the student something to evaluate rather than a finished answer to accept.
The difference can also be described through help-seeking. In a productive sequence, a learner notices a specific difficulty, asks a targeted question, evaluates the help, tries again, and adjusts the approach. Evidence reviewed by the OECD suggests that some students working with general-purpose chatbots instead ask for a solution and implement it immediately, skipping diagnosis, evaluation, and iteration. The issue is not that the chatbot answered quickly. It is that the learner lost the stages in which they discover what kind of help they need and whether the help is sound.
Ease can then become misleading feedback. When a response arrives fluently and the learner recognises its words, the interaction may feel like understanding. This is sometimes called an illusion of fluency: the experience of processing information smoothly is mistaken for the ability to retrieve or use it later. Direct-answer tools can intensify that error because they remove the visible struggle that would otherwise reveal a gap. A student may finish a reading session feeling informed, only to discover during an unaided question that the structure of the explanation was never built in memory.
The problem is therefore not solved by demanding effort everywhere. It is solved by locating the effort that matters. A calculator can be appropriate when the learning goal is modelling rather than arithmetic. An AI editor may be useful after a student has made the argument, if the student remains responsible for deciding which changes improve it. But when the target is the ability to reason, synthesise, diagnose, or explain, the learner must perform enough of that work to develop the capability.
From a missing step to a missing sense of agency
The effects of offloading can extend beyond memory. Agency in education means being able to make intentional, informed choices about one’s work: what to attempt, what to trust, what to question, and how to improve. A student who uses AI selectively and checks its contribution can remain the author of those decisions. A student who treats the system as an opaque authority may still make choices, but the choices become increasingly constrained by whatever the system presents as the obvious answer.
The pathway often begins with a reasonable motive. A student who lacks confidence may ask AI for help to avoid getting stuck. If the response solves the immediate problem, the student receives relief but less practice confronting difficulty. The next difficult problem may then feel even less manageable, increasing the appeal of asking AI to take over again. Favero and colleagues describe this as a possible self-reinforcing cycle in which avoidance of challenge weakens confidence, and weaker confidence increases dependence. Their framework is an integrative account of plausible relationships, not a new causal test of every learner, but it identifies why convenience can become a habit rather than a one-time aid.
Human-like language can make the cycle harder to notice. A chatbot presents a coherent response, adapts its tone, and can explain the same point several ways. Those features make it convenient to use, but they can also make the system seem more knowledgeable and more certain than it is. Overtrust is not simply believing a false statement. It is allowing the apparent reliability of the source to replace the learner’s own evaluation of evidence, assumptions, and uncertainty.
Once evaluation is outsourced, the student may lose more than the answer to one question. They may become less practiced at recognising a weak explanation, less willing to tolerate ambiguity, and less confident in a conclusion that does not come pre-approved. This is why the educational concern is properly described as a problem of agency as well as cognition. Independent learning requires not only information but the authority—and the willingness—to judge whether information deserves acceptance.
The broader human setting also matters. An integrative framework by Favero and colleagues links cognition, agency, emotion, and ethics rather than treating them as isolated risks. If AI-mediated work repeatedly avoids challenge, a student may experience disengagement or reduced self-efficacy. If most learning interactions become mediated by a system, opportunities for meaningful contact with teachers and peers may shrink. If the student feels that AI use conflicts with expectations of authenticity or effort, assistance can produce guilt and anxiety; if assistance becomes treated as an automatic entitlement, shared norms around authorship and fairness can weaken.
These possibilities should not be inflated into universal outcomes. The framework identifies a reinforcing risk pattern, while the experiments in this field measure particular tasks and populations. Still, it gives the performance–learning distinction a human consequence. A learner who can produce an answer but cannot explain, question, or defend it has not merely missed a test opportunity. They have less control over the knowledge they are expected to use.
Privacy and surveillance can affect that control in the opposite direction. Educational AI systems may collect prompts, drafts, behavioural patterns, and performance data. Some process evidence can help teachers understand how students learn, but a learner who believes every mistake will be stored, scored, or used to build a profile may avoid experimentation. Errors are not just failures to be eliminated; they can be information about what needs to change. A learning environment that makes mistakes feel permanently consequential can encourage compliance and polished imitation instead of inquiry.
Unequal AI literacy changes who bears these risks. Students with stronger prior preparation or digital literacy may be better able to question a response, place it in context, and use it as one input among several. Students with fewer such resources may treat a fluent system as an authority, especially when a teacher or trusted human source is not readily available. An AI tool can therefore widen differences even when it is nominally available to everyone. Access is not the same as the capacity to evaluate what access produces.
A learning-oriented scaffold responds by preserving specific learner actions. It can ask for an attempt before offering a hint. It can request an explanation of why an answer seems plausible. It can present a counterexample or a question instead of a complete solution. It can make uncertainty visible, ask the student to compare two possible arguments, and delay the polished wording until the reasoning is settled. It can also reduce support over time and include a check in which the learner must recall or apply the idea without the tool.
The OECD synthesis describes related possibilities in pedagogically configured tutors, including Socratic questioning, adaptive difficulty, clarification questions, and support that is gradually withdrawn. These features do not make a system automatically effective. They change the role the system is being asked to play. The relevant design question becomes: does the interaction elicit the learner’s reasoning, or does it merely deliver reasoning in a form the learner can repeat?
The two paths in the diagram are not claims that every interaction follows a fixed sequence. They are a way to inspect the location of cognitive work. The shortcut path can be fast and satisfying while producing weak evidence of independent capability. The scaffold path may feel slower because it keeps uncertainty and choice with the learner, but it creates opportunities for the learner to explain, evaluate, and improve.
That is the standard to carry forward: not “Did the tool reduce effort?” but “Did it reduce the effort that mattered?” The next question is whether actual users choose the scaffold path or the substitution path when they are simply given access to a general-purpose system. That distinction becomes visible when researchers compare what students do with AI, what they can do without it, and what survives after the tool is removed.