When Does AI Become Cognitive Debt? | Digital Wellness and the Thinking Students Still Need to Do

Artificial intelligence has changed one of the oldest bargains in education.

For most of school history, if a student wanted a paragraph, solution, summary, explanation, set of examples, revision plan or first draft, somebody had to do the cognitive work. The learner could copy, ask a friend, search a book or use a calculator, but the boundary between assistance and authorship was usually visible.

Generative AI makes that boundary fluid. A student can begin with a blank page and receive a polished answer in seconds. They can ask for hints, feedback, questions, explanations, alternative examples or complete solutions. The same tool can function as tutor, editor, search assistant, practice generator, translator, brainstormer, simulator or substitute thinker.

That range is why simple rules such as “AI is good” or “AI is cheating” are inadequate.

The better question is: what cognitive work is the student outsourcing, and what happens to the capability after repeated outsourcing?

September 2026 has made this question especially current. ASCD’s Educational Leadership devoted its September issue to digital wellness, including the idea that AI should support thinking rather than replace it. The Education Endowment Foundation has also announced a major research programme on generative AI and learning because the long-term effects on cognition and educational outcomes remain an open evidence problem.

A useful way to think about the risk is cognitive debt.

Financial debt gives you spending power now in exchange for obligations later. Cognitive debt gives you performance now by borrowing against learning you may still need later. The student gets the paragraph, the solution or the explanation immediately. The hidden bill arrives when they must write, solve, judge, retrieve or explain without the tool.

The point of this article is not to discourage AI. It is to help students, parents and teachers decide when AI is buying productive leverage and when it is quietly financing present performance with future dependence.

1. Cognitive offloading is normal; cognitive surrender is different

Humans have always used tools to reduce mental work.

We write lists so memory does not have to hold everything. We use dictionaries instead of memorising every definition. Calculators handle arithmetic once the relevant mathematics is understood. Maps reduce navigational memory. Spreadsheets automate repeated computation. Search engines help us locate information.

This is cognitive offloading, and it is often intelligent.

The mistake is assuming that because some offloading is useful, more offloading is always better.

Whether offloading helps depends on the learning goal. If the goal is to decide which statistical model fits a problem, using software to perform repetitive arithmetic may free attention for interpretation. If the goal is to learn multiplication facts, automating every multiplication removes the very retrieval practice the learner needs. If the goal is to produce a polished final report at work, AI editing may be efficient. If the goal is to learn to build a coherent paragraph, generating the paragraph can bypass the capability being trained.

The central question is therefore not, “Did a tool do work?” It is: Was that work part of the capability we are trying to build?

2. Performance and learning are not the same output

AI makes an old educational distinction impossible to ignore.

A student can perform well on a task without learning much from it.

A beautifully written response might reflect excellent writing capability. It might also reflect excellent prompting. A correct mathematics solution might show understanding. It might also show accurate transcription from a generated method. A fluent summary might reveal comprehension. It might also be a compressed version of a text the student never really processed.

Teachers have always faced this problem with homework help, model answers and tutoring. AI increases the scale because assistance is immediate, patient and often linguistically convincing.

This means schools need to stop treating product quality as the only evidence of learning.

We need process evidence and transfer evidence.

Can the student explain why the answer works? Can they reproduce the reasoning later? Can they identify a flaw? Can they adapt the method when the surface changes? Can they write a new paragraph without the tool? Can they decide whether the AI response is relevant, incomplete or wrong?

If not, the polished product may be masking unpaid cognitive debt.

3. The debt metaphor: what gets borrowed?

Cognitive debt can take several forms.

Retrieval debt. The tool keeps recalling facts, vocabulary, formulas or definitions that the learner never practises retrieving.

Generation debt. The tool creates sentences, hypotheses, examples or plans before the learner attempts to generate them.

Judgment debt. The tool decides which evidence matters, which method applies, what is strong or weak, or which interpretation is more plausible.

Monitoring debt. The tool checks coherence, errors and omissions so thoroughly that the learner stops learning to notice them.

Persistence debt. The tool removes every moment of being stuck, so the learner never develops strategies for uncertainty.

Representation debt. The tool converts between text, diagrams, equations and summaries before the student learns to translate representations.

Language debt. The tool continually upgrades vocabulary and sentence structure, creating prose the student cannot independently produce or even fully control.

Not every loan is bad. Experts borrow all the time. The danger appears when students repeatedly borrow the very function their education is trying to develop.

4. Productive friction is not wasted time

AI tools are often sold through the language of friction removal: fewer steps, faster answers, instant feedback, less blank-page anxiety.

Some friction should absolutely be removed. Poor interfaces, repetitive formatting, inaccessible text, needless copying and administrative clutter waste time.

But education contains productive friction.

Productive friction is the resistance created by the learning itself: trying to recall before looking, deciding which method applies, forming a sentence, comparing evidence, detecting an inconsistency, reorganising a confused explanation, or persisting long enough to discover what you do not understand.

This friction feels inefficient because learning is slower than answer production.

Yet the slowness is sometimes the mechanism.

A student who attempts a definition before checking it activates memory differently from a student who instantly requests the definition. A writer who struggles to phrase a claim discovers what they cannot yet articulate. A mathematician who chooses the wrong method learns something by diagnosing why it failed.

AI should remove accidental friction without deleting the productive friction that builds capability.

5. A strong default: human first, AI second, human last

One practical rule works across many subjects:

Human first → AI second → human last.

Human first means the learner produces an initial cognitive object: a prediction, draft, solution attempt, outline, question, explanation or evidence selection.

AI second means the tool supports the work: critique it, generate counterexamples, ask questions, suggest alternatives, provide feedback, explain a missing concept or create additional practice.

Human last means the learner regains ownership: revise, verify, reconstruct, explain, decide, retrieve or apply without copying the tool.

This sequence matters because starting with AI can anchor the student’s thinking. Once a fluent answer is visible, it becomes difficult to know what you would have generated independently. AI-first work can turn a learner into an editor of someone else’s reasoning.

Human-first work creates a baseline. It preserves evidence of the student’s current model.

Human-last work checks whether the interaction changed the learner rather than merely changing the document.

6. Writing: feedback is different from authorship

Writing is one of the clearest places to draw the boundary.

Suppose a Secondary student is writing an argumentative essay.

AI can reasonably help after an initial draft by asking, “Where is the weakest link in my reasoning?” “Which paragraph lacks evidence?” “Identify two places where the transition is unclear but do not rewrite them.” “What counterargument have I not considered?” “Ask me questions that would help me improve the conclusion.”

This use keeps the student in the author role.

The risk increases when the prompt becomes, “Write the essay,” “Improve this so it gets an A,” or “Rewrite every paragraph in a more sophisticated style.” The student may receive better prose, but the locus of generation has moved.

A useful test is edit distance from capability. If the final writing contains vocabulary, syntax, reasoning and organisation far beyond what the student can explain or recreate, the document has outrun the learner.

That may be acceptable in some professional contexts where output matters more than training. In school, it is usually a warning.

7. Mathematics: hints can preserve thinking; complete solutions can erase method selection

AI is excellent at providing worked solutions. That is both useful and dangerous.

A student stuck on algebra can ask for the full solution and feel immediate relief. But the most valuable learning question may have been: What kind of problem is this, and what should I try next?

If the AI supplies the method before the learner attempts method selection, it can remove the decision the learner needs to practise.

A better prompt ladder is:

“Do not solve it. Ask me one question that helps me identify the topic.”
“Tell me whether my first step is valid and why.”
“Give me a hint that preserves the next step.”
“Show me a simpler analogous problem.”
“Now show the worked solution.”
“Give me a fresh problem with the same underlying structure but different surface features.”

The last step matters. Transfer checks whether the student owns the method.

AI becomes educationally powerful when it controls the amount of help rather than maximising help.

8. Science: explanation must remain anchored to evidence

AI can generate plausible explanations for almost any scientific question. Plausibility is not evidence.

Students need to learn that scientific explanation is constrained. Claims must fit observations, mechanisms, data and established knowledge.

A productive AI use might be: “Here is my explanation of why the temperature changed. Identify which sentence is a claim, which is evidence and where the mechanism is missing.” Another might ask for two competing explanations and require the student to decide which better fits the data.

A weak use is asking AI to interpret the experiment before the student has examined the results.

The learner should remain the person who decides what the evidence permits.

This distinction matters beyond school science. AI systems generate fluent language. Scientific literacy requires knowing that fluency does not increase evidential weight.

9. Reading: summarisation can either deepen or bypass comprehension

Students often use AI to summarise difficult texts.

Sometimes that is a legitimate accessibility scaffold. A concise orientation can help a learner enter a dense article.

But there is a difference between using a summary to support reading and using a summary instead of reading.

If the educational goal is to understand the source, students need contact with the source. They need to resolve pronouns, track argument structure, judge evidence, notice qualifications and distinguish the author’s claim from examples.

A useful sequence is: read a section, write a one-sentence summary, compare with AI, return to the text, revise the summary, then answer a question that requires exact textual evidence.

The AI summary becomes a comparison object rather than a substitute text.

10. Verification is not a final chore; it is a central AI literacy skill

“Check the AI’s answer” sounds simple until we ask how.

To verify, a learner needs independent knowledge, reliable sources, criteria or reasoning. A student who knows almost nothing about a topic cannot confidently identify a sophisticated hallucination.

Therefore, verification itself must be taught.

Students can learn to classify AI claims:

directly verifiable from the provided source;
widely established background knowledge;
interpretation requiring argument;
current fact requiring an up-to-date source;
numerical claim requiring calculation;
citation requiring inspection;
uncertain claim that should not be repeated confidently.

They can also learn to ask, “What evidence would I need before accepting this?”

This turns scepticism into a method rather than a vague warning.

11. Fluency is a dangerous signal

Generative AI is extraordinarily good at producing language that sounds complete.

Humans often use fluency as a cue for truth and understanding. A smooth explanation feels easier to process, and ease can be mistaken for mastery.

Students therefore need to learn a counterintuitive habit: make fluent answers work harder.

Ask the AI to state assumptions. Ask for a counterexample. Ask what would falsify the claim. Compare two independent sources. Translate the explanation into a diagram. Apply it to a new case. Explain it aloud without looking.

If the idea collapses when reformulated, the student never really possessed it.

Digital wellness in the AI era is partly about resisting the seduction of frictionless language.

12. The “prompting is thinking” claim is only partly true

Prompting can involve real cognition. A sophisticated prompt may require decomposition, criteria, context, audience awareness and evaluation.

But not all prompting is educationally equivalent to doing the target task.

A student might write an excellent prompt that produces a persuasive essay without practising paragraph construction. They may specify a calculus problem accurately without learning to solve it. They may ask AI to build flashcards without retrieving the vocabulary.

Prompt skill is a capability. It should not be allowed to masquerade as every other capability.

This is similar to using a calculator. Knowing how to use the calculator is valuable. It does not prove number sense.

Schools should therefore assess prompting where prompting is the goal and assess independent subject capability where that is the goal.

13. Emotional wellness matters too

Digital wellness is not only cognitive.

AI tools can become reassurance machines. A student can repeatedly ask, “Is this good enough?” “Am I right?” “What should I do next?” The tool is always available and rarely impatient.

That convenience can support anxious learners, but it can also create dependence on external confirmation.

Students need opportunities to make decisions under uncertainty and live with provisional judgments.

A healthy AI routine may include “confidence before consultation”: the student states what they think and how confident they are before asking the tool. After the response, they note whether the evidence changed the judgment.

This preserves agency.

The goal is not to make students avoid help. It is to prevent help-seeking from becoming compulsory reassurance.

14. Social wellness: a tool should not quietly remove people from learning

AI can replace interactions that used to involve teachers, peers or family members.

Sometimes that is positive. A student may ask a question at midnight without embarrassment. A multilingual learner may rehearse language privately. A shy student may prepare before speaking.

But education is also social. Students learn to explain to people, tolerate disagreement, read confusion, negotiate meaning, ask for clarification and contribute to shared work.

If AI becomes the preferred partner for every difficult conversation, some interpersonal capabilities may receive less practice.

The useful design question is: Which interactions are better automated, and which are part of the human capability we want students to build?

15. AI should sometimes refuse to be too helpful

The ideal educational AI is not always the most compliant AI.

If a student asks for a complete answer to a practice question, the most educational response may be a diagnostic question. If they ask for a conclusion, the system might request the evidence first. If they ask for a rewrite, it might identify two weaknesses and return responsibility.

Teachers can simulate this even when the AI platform itself does not.

Give students prompt protocols:

“Do not provide the final answer until I have attempted.”

“Ask me to justify each claim.”

“Give one hint at a time.”

“After helping, test me with a new example.”

“Do not rewrite my sentence; identify the problem.”

“Require me to choose before you explain.”

These prompts make AI less efficient at answer delivery and more useful for learning.

That apparent inefficiency is the point.

16. The cognitive debt ledger

Students can use a simple ledger after AI-assisted work.

What did AI do?
Generate, explain, retrieve, calculate, critique, organise, translate, verify, brainstorm?

What did I do before AI?
Attempt, predict, draft, choose, retrieve, explain?

What changed in my understanding?
Name the concept, error, strategy or decision.

What can I now do without AI?
This is the repayment test.

What still requires the tool?
Be honest.

If a student repeatedly cannot name what became independently available, the tool may be improving products more than learners.

17. Classroom policies should be task-specific, not slogan-specific

“AI allowed” and “AI banned” are often too coarse.

Different tasks have different learning jobs.

During initial vocabulary retrieval, AI may be prohibited because recall is the exercise. During later sentence generation, AI may be allowed to produce contrasting examples. During a first essay draft, AI may be restricted. During revision, targeted feedback may be encouraged. During a research task, AI might help generate search terms but not serve as the final source. During coding, it may help explain errors after an individual attempt.

Task-specific rules make the reason visible.

Students are more likely to understand boundaries when teachers explain what capability is being protected.

The policy becomes: “AI is restricted here because this thinking is the thing you are training.”

That is educationally stronger than moral panic.

18. Assessment must evolve

If polished take-home products can no longer reliably show independent capability, assessment design needs more triangulation.

That does not mean returning every task to timed handwritten exams.

Teachers can combine:

in-class generation;
oral explanation;
process logs;
version histories;
transfer questions;
source annotations;
live problem solving;
reflective commentary;
AI-use declarations;
portfolio evidence;
short retrieval checks.

The purpose is not surveillance. It is validity.

An assessment should provide evidence about the capability it claims to measure.

If AI obscures that evidence, the design should change.

19. Parents need a more useful question than “Did you use AI?”

The binary question creates predictable behaviour. Students either say no, feel accused or hide usage.

A better conversation is:

“What did you use it for?”

“What had you done before you asked?”

“Which part of the answer did you verify?”

“What did you change because of it?”

“Can you now do the same kind of thing without it?”

“Where would using AI remove the practice you still need?”

These questions treat the student as a developing decision-maker.

Parents can also model healthy tool use. Adults use calculators, maps, spellcheck, search and AI. The lesson is not purity. It is intentionality.

Use tools where they amplify capability. Practise unaided where the capability itself matters.

20. The digital wellness test: does the tool leave the learner more capable?

A useful educational technology test can be stated in one line:

After the tool is removed, what has improved?

Possible answers include:

I can now explain the concept.

I can recognise the error.

I can write the next paragraph better.

I can select the method.

I can ask a sharper question.

I can verify a claim.

I can solve a similar problem.

I can plan the task.

I can notice when I need help.

Those are signs of productive leverage.

Warning signs include:

I can submit the work, but I cannot explain it.

I need the tool to start every time.

My vocabulary in the document is not vocabulary I use.

I accept answers because they sound right.

I have stopped checking sources.

I feel unable to decide without asking.

I understand while reading the AI response but cannot reconstruct it later.

Those are signs the debt may be accumulating.

21. A worked English case: feedback without replacement

A student writes a situational response.

Before AI, they identify audience, purpose and three required content points. They draft the first version.

Then they ask AI: “Do not rewrite. Identify one place where the tone does not fit the audience, one missing content point and one sentence whose meaning is unclear. Explain why.”

The student reviews the feedback and decides which suggestions are valid. They revise.

Finally, they close the tool and write a new opening for a similar situation from scratch.

This sequence uses AI to increase feedback frequency without transferring authorship.

The final independent task repays the debt.

22. A worked mathematics case: from rescue to diagnostic tutoring

A student is stuck on a quadratic equation.

The unhelpful route is: “Solve this.”

The educational route begins with: “I think this is factorisation, but I am not sure. Do not solve it. Tell me what feature should determine whether factorisation is sensible.”

The AI responds with criteria. The student attempts.

If the attempt fails, they ask: “Point to the first invalid step only.”

After repair, the student asks for a fresh quadratic with the same structural idea.

They solve that independently.

The AI has functioned as a tutor because the student retained method selection, execution and transfer.

23. A worked research case: AI as map, sources as ground

A student begins a research project on urban heat.

AI can help generate subquestions: surface materials, vegetation, building density, transport, health effects, measurement methods.

That is useful orientation.

But the student should then move to reliable sources. AI does not become evidence merely because it generated a bibliography-shaped answer.

The student verifies claims in government data, scientific literature and trusted institutional sources. They use AI later to compare their outline against the research question or identify missing counterarguments.

This division of labour is powerful:

AI helps map the territory.

Sources establish the ground.

The student makes the judgment.

24. Digital wellness is not anti-technology; it is pro-agency

The phrase digital wellness can sound like a polite way of saying “less screen time.”

That is too narrow for AI.

A student can spend only twenty minutes with an AI tool and still outsource a crucial act of thinking. Another can spend an hour using technology to simulate, practise, compare, create and receive feedback while remaining intellectually active.

Time matters, but function matters more.

Digital wellness should include cognitive agency: the ability to decide when a tool is helping, when it is replacing, when to verify, when to work unaided and when to stop.

That is a twenty-first-century learning skill.

25. The long-term objective: students who can use powerful tools without becoming powerless without them

Education has never been about rejecting tools.

Books are tools. Algebraic notation is a tool. Graphs are tools. Calculators are tools. Search is a tool. AI is an unusually flexible tool.

The educational challenge is to make sure the learner develops alongside the tool.

A strong graduate should be able to use AI to extend reach, increase feedback, test ideas, accelerate routine work and access explanations. They should also be able to recognise when the task requires unaided retrieval, original generation, human conversation, disciplined reading, manual calculation or independent judgment.

They should not confuse assistance with mastery.

They should know that convenience has an opportunity cost.

They should be able to ask the most important question of any intelligent tool:

What thinking am I still responsible for?

Conclusion: repay the cognitive debt before it comes due

AI can give students extraordinary leverage.

It can make expertise more accessible, reduce some barriers, multiply examples, personalise practice, provide rapid feedback and help learners ask questions they might never ask aloud.

The danger is not that the tool is powerful.

The danger is that students can receive the appearance of capability before capability has formed.

That is cognitive debt.

The solution is not blanket prohibition. It is disciplined sequencing: human first, AI second, human last. Protect productive friction. Make students verify. Distinguish product from learning. Design transfer checks. Teach task-specific boundaries. Require students to notice what became independently available.

Then AI becomes something better than a shortcut.

It becomes a lever that leaves the learner stronger after the lever is removed.

Research basis and further reading

ASCD Educational Leadership, Keeping Students Digitally Healthy, September 2026 issue.
ASCD, Shifting the Digital Wellness Conversation (1 September 2026).
ASCD, Teaching Students to Use AI for Writing Feedback (1 March 2026).
Education Endowment Foundation, New landmark programme of research to investigate the impact of Generative AI on learning and cognition (8 June 2026).
Edutopia, 5 Tech Strategies to Enhance Student-Led Learning (12 March 2026).
eduKateSG collision review included How Education Works | AI Literacy Education, How Studying Works | Cognitive Offloading, How to Think Properly | Use AI Without Handing Over the Final Judgement, How Studying Works | The Locus of Generation, and the new AI Self-Regulation Scaffolding article. This article’s canonical job is narrower: digital wellness as the management of cognitive debt across retrieval, generation, judgment, monitoring and independent transfer.

26. The capability floor: some things should remain available without the tool

A useful way to think about AI policy is to distinguish between a learner’s capability floor and capability ceiling.

The ceiling is what the learner can accomplish with excellent tools, good prompts, strong references, collaboration and enough time. AI can raise that ceiling dramatically. A student may analyse more examples, explore alternative structures, test more scenarios or receive richer feedback than would otherwise be practical.

The floor is what remains available when support is reduced.

Education needs both.

If the ceiling rises while the floor collapses, the learner becomes impressive under ideal conditions and fragile under ordinary ones. They may produce sophisticated work but struggle when the network fails, the examination is closed-book, the prompt is unfamiliar, the AI answer is wrong, or a real conversation requires an immediate judgment.

A sensible curriculum therefore identifies capabilities that should remain personally available. The list differs by age and subject, but it might include foundational vocabulary, basic number relationships, sentence construction, core scientific concepts, source evaluation, common algebraic transformations, reading comprehension, note-making, planning and the ability to explain a decision.

This is not nostalgia for tool-free education. Pilots use instruments but still train for failures. Doctors use diagnostic systems but still require foundational knowledge. Engineers use software but still need enough understanding to recognise impossible outputs.

The capability floor gives the learner a platform from which tools can be used intelligently.

Teachers can protect the floor with low-stakes unaided moments: short retrieval, oral explanation, quick estimation, handwritten planning, mental checks, independent summaries and transfer questions. These should not dominate every lesson. They simply keep essential functions alive.

The principle is straightforward: use AI to raise the ceiling, but do not let it quietly lower the floor.

27. The recovery drill: what happens when the AI answer is wrong?

One of the best ways to teach digital wellness is to deliberately use imperfect AI output.

Give students a response that is fluent but contains one wrong assumption, one unsupported claim or one subtle mathematical error. Do not tell them where the problem is. Ask them to audit it.

This changes the learner’s relationship to the tool. AI stops being an oracle and becomes an object of analysis.

Students can learn a recovery routine:

Pause. Do not repair by asking the AI to “fix itself” immediately.

Locate. Identify the exact sentence, step or claim that appears doubtful.

Classify. Is the issue factual, logical, mathematical, evidential, linguistic or contextual?

Ground. Return to the original source, rule, data or calculation.

Repair. Produce the corrected reasoning in the student’s own words.

Retest. Apply the corrected principle to a new case.

This routine is valuable because real-world AI use will not always announce its errors. The user must notice.

The deeper lesson is that verification is not something students do after using AI. Verification is part of competent AI use from the beginning. The more fluent the system becomes, the more important independent judgment becomes.

That may be the central paradox of AI literacy: as tools become easier to trust, learners need stronger reasons for knowing when trust is deserved.