Cognitive Offloading: What Should We Hand Over to AI?
In the first post of this series, I focused on prompting as a form of thinking. Before students can clearly tell AI what they need, they have to define the problem, identify relevant context, recognize constraints, and decide what kind of support would be useful.
But once a student decides what to ask, another question follows. What part of the thinking should they hand over to AI?
This brings us to cognitive offloading.
Cognitive offloading is the use of an external action, object, or tool to reduce the mental demands of a task. Psychologists Evan Risko and Sam Gilbert describe it as changing how a task must be processed so that less work has to be performed internally.
The idea did not begin with generative AI.
We write appointments on calendars, so we do not have to remember them. We create shopping lists, use calculators, set reminders, take photographs, record notes, and rely on GPS. Each of these tools allows us to move part of a cognitive task into the environment.
In many cases, that is useful.
Our working memory is limited. We cannot hold every detail in mind at once, and attempting to do so can interfere with more important thinking. Cognitive offloading can free attention for interpretation, planning, problem-solving, and decision-making. It can reduce unnecessary cognitive demands and help us manage complex tasks more effectively.
The question, then, is not whether students should ever offload thinking. They already do. We all do. The more important question is what they are offloading and why.
Generative AI changes this question because it can do far more than store information or perform calculations. It can summarize readings, generate explanations, organize ideas, construct arguments, recommend solutions, interpret evidence, produce drafts, and make decisions.
AI does not simply help us remember where we put the grocery list. It can participate in forms of thinking that education is specifically intended to develop.
Research on cognitive offloading suggests that people make decisions about whether to rely on external support based partly on how difficult they believe the task will be and how confident they are in their own ability. But those judgments are not always accurate. We may offload because we underestimate our ability, overestimate the difficulty, or simply want to avoid mental effort.
That last point may sound negative, but it is deeply human. Our minds naturally look for efficient ways to manage cognitive demands. If a tool can make a task easier, using it may feel like an entirely rational choice.
Students are no different.
If AI can summarize the reading, generate the discussion post, solve the problem, or organize the assignment, why would a student choose the more difficult path?
The answer depends on what the difficulty is doing.
Some difficulty is unnecessary. A student should not have to use limited mental resources remembering complicated formatting instructions when those resources could be used to develop an argument. A student with a documented learning need may benefit from AI helping make complex language more accessible. A student beginning an unfamiliar task may need an example before they can make sense of the process.
Reducing those demands can support learning. But some forms of difficulty are part of how learning happens.
Cognitive psychology has repeatedly demonstrated that conditions that make learning feel easier do not always produce stronger long-term learning. Robert and Elizabeth Bjork use the term “desirable difficulties” to describe learning conditions that may slow performance in the moment but strengthen retention and transfer over time.
Retrieving information, generating an answer, comparing possibilities, working through uncertainty, and correcting mistakes can all require effort. That effort may be uncomfortable, but it is often where knowledge becomes more durable.
The generation effect offers a simple example. Research dating back to Norman Slamecka and Peter Graf found that people tend to remember information better when they generate it themselves than when they simply read information someone else has provided.
That does not mean students should discover every concept independently or struggle without guidance. Cognitive load research also shows that novices can be overwhelmed when they are asked to solve complex problems without adequate knowledge or support. Sometimes an explanation, worked example, hint, or scaffold is exactly what the learner needs.
The challenge is distinguishing between the cognitive effort that interferes with learning and the cognitive effort that creates it.
AI can reduce both.
It can remove unnecessary friction, but it can also remove the generation, retrieval, evaluation, and decision-making students needed to practice.
Imagine a student who uses AI to organize notes from several sources. If the learning goal is to construct an argument using those sources, that organizational support may free the student to focus on interpretation and judgment.
But imagine that the student asks AI to identify the main ideas, decide how the sources relate, generate the argument, and organize the evidence. The student has not simply offloaded organization. They may have offloaded the sensemaking the assignment was designed to develop.
The final product may look stronger. The student may complete the task faster. But a better product does not necessarily mean more learning occurred.
This distinction between performance and learning is especially important in an AI-rich environment.
Performance describes what students can accomplish under the current conditions, including the tools and support available to them. Learning describes what has changed in the student—what they understand, remember, can explain, and can apply when the support changes or disappears.
AI can improve performance while it is present without necessarily building the capability students will need when it is absent.
A 2025 field experiment involving nearly 1,000 high-school mathematics students illustrates this tension. Researchers Hamsa Bastani and colleagues compared students working without AI, students with access to a standard GPT-4 interface, and students using a more carefully designed AI tutor.
Students with unrestricted GPT-4 access performed substantially better during practice. But when the AI was removed for an assessment, their performance fell below that of students who had practiced without it. The tool had helped them complete the immediate problems without necessarily helping them develop the same ability to solve problems independently.
However, the study also found something important: the outcome changed when the AI was designed with guardrails. The more structured tutor was instructed to provide hints and guidance without simply supplying complete solutions. Students still received support, but the support was less likely to replace the problem-solving process.
That difference reinforces a central point. The effects of AI on learning do not come from access alone. They depend on the role AI is given and how students interact with it.
More recent research points in the same direction. In a 2026 randomized study of undergraduates, Zara Contractor and Germán Reyes found that students with access to generative AI showed gains on immediate and delayed knowledge assessments. But those gains were stronger among students who used AI for augmentation—such as requesting explanations—than among students who used it for automation, such as generating text for them.
These findings should make us cautious about broad claims that AI either improves learning or damages it. AI is not one kind of learning intervention.
Asking AI to explain a confusing concept is not cognitively equivalent to asking it to write the assignment. Requesting a hint is not the same as requesting the solution. Using AI to organize information is not the same as allowing it to interpret the information and make the final decision.
The details matter. The student’s purpose matters. The timing matters. Most importantly, the location of the thinking matters. This is why students need more than policies telling them whether AI is permitted. They need help recognizing what cognitive work a task is intended to develop.
Without that understanding, students may evaluate AI use according to only one question: Will this help me complete the assignment?
From that perspective, offloading as much as possible may seem sensible. But if the purpose is learning, students need to ask different questions. What ability is this task helping me develop? What part of the process do I need to practice? What can AI help me do without removing that practice? What might I be unable to do later if AI does this for me now? Those questions turn cognitive offloading into a metacognitive decision.
A student studying for an exam might ask AI to generate practice questions. That offloads the work of creating examples while preserving the student’s work of retrieving and applying knowledge. But if the student asks AI for both the questions and the answers before attempting them, the tool may remove much of the retrieval practice.
A student writing a paper might ask AI to format citations. If citation formatting is not the central learning goal, that may be a reasonable use of the tool. But if the student asks AI to select the evidence, interpret the sources, and decide what argument they support, the student may be offloading the very judgment the assignment was intended to develop.
A student troubleshooting a technical problem might use AI to create a checklist of possible causes. That could expand the range of possibilities the student considers. But if the student follows the checklist without understanding why each step matters, they may complete the repair without developing troubleshooting ability.
In each case, AI can support the work or substitute for it. The difference is not always visible in the final product. That is why I think students could benefit from conducting a simple effort audit when they use AI.
· What did I do?
· What did AI do?
· What decisions remained mine?
· What part of the task required me to retrieve, generate, interpret, evaluate, or apply knowledge?
· Could I explain or perform this without the tool?
The purpose of an effort audit is not to calculate an acceptable percentage of human versus AI work. Learning cannot be measured that neatly. The purpose is awareness.
Students need to notice when AI reduces an unnecessary burden and when it removes an opportunity to develop capability. They need to recognize that something feeling easier is not the same as having learned it. They need to distinguish between completing a task with assistance and becoming more capable because of that assistance.
Faculty have a role here as well. If we want students to make thoughtful decisions about cognitive offloading, we have to make the purpose of the work visible. Students need to know why they are being asked to generate an explanation, work through a problem, compare alternatives, or justify a choice.
Without that clarity, warnings about overreliance may sound like demands to perform unnecessary work simply because that is how the task has always been done.
The goal is not to protect every form of effort. Some effort should be reduced. Some tasks should become easier. Some cognitive work can be delegated so that students have more capacity for something more important.
But we need to be intentional about what replaces it. If AI summarizes the information, what will the student do with that summary? If AI generates possible solutions, what criteria will the student use to evaluate them? If AI creates a first draft, what intellectual work will the student perform during revision? If AI explains the concept, how will the student demonstrate that the explanation became understanding? The question is not simply whether AI did part of the work. The question is what happened to the student’s thinking because it did.
Cognitive offloading is not automatically a threat to learning. Used thoughtfully, it can help students manage complexity, access support, and devote more attention to higher-level thinking. But when students offload the thinking, they still need to develop, efficiency can become dependency.
That is the tension at the center of learning alongside AI.
Students need to learn what they can responsibly hand over, what they should keep, and how to tell the difference. Because the goal is not to ensure that students always work without assistance.
The goal is to ensure that the assistance helps them become more capable, not merely more productive. Once students decide what role AI should play, they face another challenge. They have to determine whether the information AI provides is accurate, complete, and appropriate for the context.
That will be the focus of the next post: why verification requires much more than checking a few facts.
Continuing the Conversation
Series 1: AI Is Exposing Existing Problems ✓ Completed
Series 2: What We Do About It ✓ Completed
Series 3: Cultivating Human Thinking ✓ Completed
Series 4: Learning Alongside AI
Current Post (2 of 8): Cognitive Offloading: What Should We Hand Over to AI?
Next Up: Verification Is More Than Fact-Checking