A student is asked to explain why a proposed solution to a mathematics problem fails. She opens a chatbot, pastes the problem, and receives a lucid explanation in seconds. It identifies the false assumption, names the correct principle, and provides an elegant alternative. The explanation is better organized than anything she might have produced herself.
She submits it. Her assignment is complete.
But what, exactly, has she learned?
The question is not whether the chatbot was useful. It plainly was. The question is whether the student acquired the capacity to notice the same error tomorrow, when the screen is closed and the problem looks different. We now have technologies that can deliver the appearance of intellectual accomplishment before the learner has performed much of the intellectual work. Education has not yet settled on a way to distinguish the two.
AI did not invent intellectual shortcuts. Libraries, calculators, search engines and collaborative writing have always changed what people needed to remember or do alone. What is new is the breadth of the work that generative AI can perform: it can produce the explanation, draft the argument, correct the answer, imitate the skeptical examiner, and propose the next question. It can enter almost every stage of thinking, not just its final calculation.
The central educational dilemma of the AI era is therefore not whether we should outsource thinking. We already do. It is which thinking we can safely outsource, at what point, and for what purpose.
The important distinction: completing the task is not learning the task
Imagine two students who submit equally strong essays on a subject they barely knew the week before. One spent the afternoon searching, reading, drafting, revising and arguing with a chatbot about the evidence. The other asked the chatbot to produce the essay and then polished its tone. Both products may meet the rubric. Yet the processes might leave them with very different abilities to explain the subject without help.
This difference is easy to overlook because schools generally see finished products. A paragraph can be marked; the internal sequence of false starts, decisions and revisions that made it possible is harder to observe. A fluent answer becomes a tempting proxy for understanding.
We should distinguish at least three forms of intellectual work:
Finding and handling information. Locating facts, retrieving formulas, organizing references, translating vocabulary and keeping track of details.
Constructing understanding. Generating an explanation, connecting ideas, attempting a solution, noticing contradictions, and changing an inadequate model of the problem.
Exercising judgment. Deciding what is credible, relevant, fair, uncertain or worth doing; taking responsibility for a conclusion and its consequences.
AI can assist with all three. But not all three have the same educational status in every lesson. If the purpose is to learn a new language, having a model translate every sentence removes an important practice opportunity. If the purpose is to evaluate evidence about a historical controversy, using translation to access a source may expand the learner's opportunities to think. The same tool can obstruct one learning goal and enable another.
The educational question must come before the prompt: what ability is this activity supposed to build?
Outsourcing cognition is normal. Outsourcing practice is different.
Psychologists Evan Risko and Sam Gilbert use the term cognitive offloading for actions that reduce the mental demands of a task. Writing a reminder, rearranging objects, or using a device to retain information can free resources for other work. Their 2016 review makes an important point: offloading is a familiar feature of ordinary cognition, not an unprecedented failure of willpower. Read the review.
An architect does not need to memorize every building regulation to exercise professional judgment. A scientist can use software to analyze data without abandoning scientific reasoning. An adult can set a calendar reminder and still take responsibility for keeping an appointment. The point of learning is not to reconstruct civilization from memory.
But there is a crucial difference between using a tool after developing competence and letting a tool prevent that competence from forming.
Consider a musician practicing a difficult passage. Listening to a flawless recording is valuable: it provides a model, reveals phrasing and can correct misinterpretation. Yet listening is not the same as acquiring the motor and perceptual coordination needed to perform. An AI-generated argument can play a similar role. It may be an excellent example without being evidence that the reader can argue.
The dividing line is not technology versus purity. It is support versus substitution. A support changes the conditions in which the learner practices. A substitution performs the targeted practice in the learner's place.
What the experiments actually tell us
The evidence is more interesting than either fashionable extreme—“AI makes students stupid” or “AI makes everyone smarter.”
When performance rises but independent learning falls
In a field experiment involving nearly 1,000 high-school mathematics students in Turkey, Hamsa Bastani and colleagues compared unassisted practice with two kinds of GPT-4-based support. One interface resembled a general chatbot; the other incorporated teacher-designed safeguards that guided students with hints rather than readily supplying final answers.
Students using AI performed better during supported practice. But in the subsequent unassisted examination, the group with the general-purpose interface performed worse than the no-AI comparison group. The researchers reported a 17% reduction in grades relative to the control group for that condition. The safeguarded tutor largely avoided this disadvantage. Study in PNAS, 2025.
This finding should trouble anyone who measures an educational technology only by the scores it produces while students are using it. Better-looking work during a lesson is not necessarily stronger independent capability afterward.
Yet its limits matter. The study concerned particular mathematics tasks, students, interfaces and a short evaluation period. It does not establish that every use of a chatbot weakens every kind of learning. What it demonstrates is a mechanism that educators should take seriously: a system can make practice easier while depriving the learner of practice.
When carefully designed AI tutoring helps
A separate randomized trial published in Scientific Reports in 2025 compared a purpose-built AI physics tutor with in-class active learning in a university setting. Students using the tutor showed higher measured learning gains while typically spending less time on the material. The tutor was not merely an unrestricted answer machine: its design incorporated structured explanations, interactions and feedback grounded in pedagogical principles. Kestin and colleagues, 2025.
That result deserves as much attention as the negative one. It suggests that generative AI can improve instruction when the system is designed around the learner's process rather than simply the learner's request for a finished answer.
But this was also a specific implementation and instructional context. It cannot be converted into a general claim that an arbitrary chatbot is superior to a human teacher, or that all learning outcomes improve indefinitely. The comparison tells us something about design, not about an inevitable technological destiny.
A newer complication: what students ask AI to do
In a July 2026 IZA discussion paper, Zara Contractor and Germán Reyes describe a randomized experiment in which undergraduates learned an unfamiliar topic and wrote an analytical essay, with or without access to an off-the-shelf generative AI tool. Students completed unaided assessments immediately and one week later. The authors found gains in test performance among students given AI access, with improvements persisting over the short follow-up period.
An especially suggestive distinction emerged between augmentation—using AI to explain, question or help the student's work—and automation—using it to produce work on the student's behalf. The more durable benefits were associated with augmentation; some short-term writing advantages associated with automation did not persist after AI was removed. Contractor and Reyes, IZA Discussion Paper 18792.
This is a working paper, not a universal verdict, and its follow-up spans only a week. Even so, the three studies together suggest a better question than whether AI is good or bad for education: what kind of encounter with AI produces independent capability later?
The seductive danger of a beautiful explanation
The most dangerous educational illusion is not necessarily a wrong answer. It is a correct answer that makes us feel we understand something we cannot reproduce, use or challenge.
A student asks for an explanation of natural selection. The model writes an elegant response with examples, analogies and a confident conclusion. Everything sounds familiar. The student nods. Ten minutes later, asked to explain how natural selection differs from an organism deliberately adapting, they hesitate. They recognized an explanation; they did not yet construct one.
This distinction predates generative AI. In classic experiments published in 2006, Henry Roediger and Jeffrey Karpicke found that retrieving material through testing improved later retention compared with additional study, even though repeated exposure could create a strong feeling of knowing. Test-Enhanced Learning, 2006.
Retrieval is not the only route to understanding. Some complex tasks demand worked examples, instruction, collaboration and extended guided practice. But the research helps expose why a tool that makes information continuously visible can distort the learner's sense of mastery. If an answer appears whenever one begins to struggle, the learner may never discover what they can actually summon and organize alone.
In this sense, AI can create fluency without ownership: the intellectual content passes smoothly before the learner, but the learner has not yet taken responsibility for it.
An elegant explanation is a beginning. It is not a certificate of mastery.
The expertise paradox
The people who most need an explanation may be least equipped to judge it.
An expert lawyer using an AI system to summarize a document can identify missing authorities, challenge questionable interpretations and notice which issue requires human scrutiny. A novice law student may not see the omissions. The expert's accumulated knowledge makes the shortcut safer; the novice's missing knowledge makes the same shortcut more tempting and less reliable.
This produces a paradox for education. Beginners may gain enormously from immediate feedback, simplified explanations and patient repetition. They may also be vulnerable to accepting unsupported assertions because the language sounds authoritative. A sophisticated interface does not turn a novice into an expert simply by placing expert-like prose on the screen.
The answer is not to forbid beginners access to useful support. It is to make support more structured at the point where judgment is weakest.
For a novice, a valuable tutor might ask what they already know, provide a carefully chosen worked example, request a first attempt, offer a narrow hint and then check the learner's explanation. For a more experienced student, the same system might be most valuable as a demanding critic: test the assumptions, present rival interpretations, and uncover counterexamples.
Scaffolding should change with competence. If it never fades, guidance can harden into dependency.
The interval that education must protect
There is a small, easily erased interval between encountering a problem and receiving a solution. During that interval, the learner predicts, recalls, selects, fails, revises and sometimes discovers that the original question was poorly framed.
It would be foolish to romanticize all struggle. Repeated confusion without guidance can waste time, lower confidence and entrench errors. A novice facing a problem far beyond their current knowledge may need an explanation before any meaningful attempt is possible. The goal is not maximum difficulty. It is an appropriate amount of cognitive work with feedback.
Still, that interval deserves protection. Once a ready-made solution arrives, the opportunity to make an unaided first attempt may vanish.
Consider a history assignment asking why two historians disagree about the causes of a revolution. A student could ask AI for a polished comparison. Or they could first identify one disagreement, examine the evidence, and write a tentative interpretation. Only then might they ask the model to produce a counterargument or identify evidence they overlooked.
The second path may be slower. That extra time is not automatically inefficiency. If the purpose is to learn historical reasoning, it may be the point of the assignment.
A learning environment should not simply minimize the time between question and answer. It should optimize what the learner becomes capable of doing in that time.
A practical rule: learner first, AI second, learner again
One useful protocol for students, teachers and adult learners can be summarized in five movements. It is a design proposal informed by the research above, not a scientifically validated universal formula.
1. Name the capability. Before opening AI, complete the sentence: “When this session ends, I want to be able to…” The answer might be to solve quadratic equations, identify a weak argument, explain a theory, or plan a credible research method. Without a target ability, convenience becomes the default objective.
2. Make a first move. Try a small piece of the work unaided: outline a claim, solve the first step, recall the key concepts, or state what is confusing. The attempt need not be heroic or lengthy; it must be real enough to reveal the next learning need.
3. Request help rather than replacement. Ask the system to offer a clue, compare two approaches, identify the first mistaken step, generate an example or challenge a tentative answer. An effective request might read: “Do not solve this yet. Ask me one diagnostic question, then give me the smallest useful hint.”
4. Rebuild without the screen. Close the answer. Explain the idea from memory, solve a fresh problem, or produce a new argument. Compare with a trustworthy source or qualified teacher, and correct errors. The student needs an independent check—not merely the pleasant sensation that the explanation was clear.
5. Return after a delay. Try again later, preferably in a changed context. If the ability survives without assistance, there is stronger evidence of learning than a polished output produced with the tool still open.
This protocol does not require every learner to perform every task without accommodations. Accessibility tools, language support and assistive technology may be essential to enable participation. The principle is to preserve practice in the skill actually being developed, not to impose the same obstacles on everyone.
The assessment problem: whose competence are we measuring?
Generative AI has exposed a weakness in assessment that existed before chatbots: the assumption that a finished artifact reveals how it was made.
If a course values the ability to research with modern tools, assess that openly. Ask students to document decisions, evaluate sources, identify model errors and defend what they chose to keep or reject. AI-assisted work can demonstrate real competence when the student owns the inquiry.
If a course values independent reasoning, it also needs moments when students reason independently. A brief oral explanation, a fresh problem under appropriate conditions, a supervised writing exercise or a discussion of an unseen counterexample may reveal capabilities a polished take-home essay cannot.
The two forms of assessment should not be confused. “Unaided” does not mean morally superior in every context; “AI-assisted” does not mean academically empty. Each measures different things.
Schools also need to acknowledge an equity problem. Students do not have equal access to reliable devices, paid systems, stable internet connections, supportive adults or time. Nor do learners share the same language background, disability-related needs or prior knowledge. An assessment that quietly assumes premium AI access may reward resources rather than learning. A blanket prohibition may remove legitimate supports without solving the underlying assessment problem.
Fairness requires explicit rules, access where appropriate, and a clear statement of what is being evaluated.
Teachers cannot outsource educational judgment
It is possible to imagine an educational system where AI generates lessons, marks assignments, proposes interventions and sends individualized messages to families. Some of these tasks may be made more efficient. None eliminates the need to ask whether the curriculum is worth learning, whether a student is being treated fairly, or whether a particular difficulty calls for challenge, reassurance, direct instruction or another kind of support.
Those are not simply information-processing tasks. They involve relationships, context, values and accountability.
A teacher who uses AI to produce three differentiated examples but watches carefully which misconception each reveals may be exercising more professional judgment, not less. A teacher who accepts AI-generated feedback wholesale because it is articulate may be surrendering the very expertise education needs.
The challenge for learning coaching is similar. The objective is not to make the coach a better answer distributor. It is to help the learner observe their own strategies, recognize the next useful difficulty, choose a form of support and gradually assume greater responsibility.
The strongest human contribution may increasingly be the ability to decide when not to provide the answer—and how to make that decision humane rather than punitive.
Where AI really can expand the mind
There is another mistake to avoid: treating every saved minute as an intellectual loss.
For a student reading a scientific article in a second language, a careful translation can open a field of thought previously inaccessible. For a teacher managing an overcrowded class, generating several appropriately pitched examples may leave more time for dialogue and individual feedback. For an adult returning to study after years away, a patient explanation at midnight may make continued learning possible.
In such cases, AI is not necessarily replacing the learner's thinking. It can be removing an obstacle that was preventing that thinking from happening.
The distinction turns on the purpose. If the goal is to learn the second language, translation should not always substitute for comprehension practice. If the goal is to evaluate a scientific argument, translation may be a legitimate scaffold. The same intervention can be appropriate in one task and inappropriate in another.
What matters is whether the tool opens a path to deeper engagement or closes the work at the first plausible answer.
The larger ambition should therefore be neither a school without AI nor a school in which AI performs every demanding task. It should be a school that knows what to automate, what to assist, what to practice, and what must remain a human decision.
Two books at the edge of this debate
Learning Coach Editions approaches the subject from a declared editorial position: learning should lead toward greater autonomy, not merely better-looking schoolwork.
The English-language book You Were Never Taught How to Learn addresses study habits, the illusion of understanding, independent learning and the difference between using AI as a support and letting it do the work. How the Brain Learns by Solving Problems explores the role of attempted solutions, feedback and correction in developing understanding. Both are listed in the English Learning Coach books catalogue.
These titles are relevant intellectual companions, not independent scientific proof of this article's claims. The empirical arguments here rest on the separately identified research. Readers interested in practical study methods can also explore the Learning Coach article Active Recall for Beginners.
The question worth keeping
A century ago, people worried that easy access to books would weaken memory; later, calculators and the internet provoked related anxieties. The answer was rarely to reject the tools entirely. It was to reconsider which abilities remained valuable and how those abilities should be taught.
Generative AI forces that reconsideration across a far wider field of intellectual activity. It can make explanations abundant, criticism immediate and drafts cheap. That is an extraordinary opportunity. It is also an invitation to confuse having access to thought with having developed the capacity to think.
The most consequential measure of educational AI may not be the quality of what a student can produce with it. It may be what the student can understand, judge and attempt after the assistance is gone—and what new questions they are capable of asking when it returns.
An answer can finish a task. Education must still build the person who knows what to do with the answer.
Research and reading notes
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). “Generative AI without guardrails can harm learning: Evidence from high school mathematics.” Proceedings of the National Academy of Sciences, 122(26), e2422633122. DOI. Randomized field study in a specific high-school mathematics context.
Kestin, G., Miller, K., Klales, A., Milbourne, T., et al. (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. DOI. Trial of a specifically designed university physics tutor.
Contractor, Z., & Reyes, G. (2026). “Experimental Evidence on the Learning Impact of Generative AI.” IZA Discussion Paper No. 18792. Working paper. Findings are preliminary and include a one-week follow-up.
Risko, E. F., & Gilbert, S. J. (2016). “Cognitive Offloading.” Trends in Cognitive Sciences, 20(9), 676–688. DOI. Review of cognitive offloading and metacognition.
Roediger, H. L., & Karpicke, J. D. (2006). “Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention.” Psychological Science, 17(3), 249–255. DOI. Experimental evidence on retrieval and later retention.
