Should Kids Use AI for School? Protect the Friction

Kids can use AI for school under one non-negotiable condition: the struggle stays intact. The reason is not nostalgia for hard homework; it is how minds get built. A child’s cognition is constructed through effortful work, retrieval that strains, drafts that fail, answers generated before solutions are seen, and an answer-on-demand machine deletes precisely that construction work while appearing to help with it. The Build First Brain rule for minors is sequencing: attempt first, AI second, always, with the model cast as a questioning tutor and feedback source rather than an oracle. The stakes differ from the adult debate in kind, not degree: an adult who outsources thinking loses practice on a built faculty. A child who outsources it during the building years may never pour the foundation at all.

Why is friction the active ingredient of learning?

Because the difficulty is the mechanism, not the obstacle. The learning-science verdict has been stable for decades: desirable difficulties, conditions that make practice harder, like spacing, interleaving, and testing yourself before seeing answers, slow performance today and massively improve retention tomorrow. The flip side condemns the easy path: the major review of study techniques ranks effortful retrieval at the top and fluent rereading near the bottom, and the generation effect shows material a learner produces themselves is remembered far better than material merely received.

Now place a frontier model in a twelve-year-old’s homework loop. Every one of those mechanisms, the strain of recall, the failed first draft, the self-generated answer, is exactly what the model offers to do instead. The tool is not cheating the assignment; it is cheating the construction. The essay still gets written. The writer does not get built.

Where AI enters the loop	What the child practices	Developmental effect	Verdict
After a genuine attempt, as feedback	Retrieval, drafting, then revision against critique	Amplifies the workout	Best use
As a Socratic questioner	Explaining, defending, re-deriving	Strengthens understanding	Excellent
As a first-resort answer engine	Prompting and pasting	Deletes the constructive struggle	The harm case
Unsupervised, ambient, always-on	Dependence as a default posture	Never learns what unaided feels like	Worst case

What makes minors a special case?

The asymmetry between losing and never-having. Adult cognitive offloading is a real but recoverable problem, faculties going frail through disuse, reversible with practice. Childhood is when the faculties are assembled: the circuits for sustained attention, working through frustration, holding an argument in the head, are wired by years of doing exactly those things, and the window is not infinitely reopenable. This is the strongest version of the argument running through why AI tutors will ruin your child’s mind: the answer-machine pattern, applied during construction, does not merely skip homework, it skips the homework’s entire developmental purpose, while producing grades that hide the vacancy, the illusion of competence in its purest form.

The same asymmetry explains why this is a protection issue and not just a pedagogy preference, the framing the neuro-rights conversation applies to developing minds generally: a child cannot consent to a dependency whose costs arrive in adulthood, so the adults run the boundary, exactly as covered for screens broadly in screen-free parenting as a competitive advantage.

What does good use actually look like?

Like a tutor who asks instead of answers, arriving after the attempt. The sequencing rule does most of the work: homework gets a genuine unaided try, the child marks where they got stuck, and only then does the model enter, to explain the stuck point, generate practice problems, or critique the draft. Used as a Socratic partner, why do you think that, what would happen if, defend your answer, the model can deliver something genuinely new in education: infinitely patient questioning calibrated to the child, which strengthens precisely the muscles answer-mode atrophies. The closing ritual seals it: the child explains the final understanding aloud, in their own words, the oral defense that no model can sit on their behalf. The mistake I see most often in households is policing the tool instead of the sequence, the issue was never whether the model was open, but whether the attempt happened first.

What should schools and policy do?

Mediate, gate, and teach the tool explicitly. The international baseline already points this direction: UNESCO’s guidance on generative AI in education calls for teacher-mediated use, age thresholds, data protection for minors, and explicit AI-literacy instruction rather than bans or laissez-faire. The implementation details that matter most track the developmental logic: assessment redesigned to reward visible reasoning, oral components that make outsourcing detectable, in-class unaided work that keeps the unassisted baseline trained, and AI literacy as curriculum, since the children will live among these systems and deserve to understand them as machinery rather than magic. Prohibition fails on contact, it relocates use to unsupervised bedrooms; surrender fails slower and worse. The school’s job is the same as the parent’s, run the boundary the child cannot yet run.

When does the protective instinct overshoot?

When it freezes into pure abstinence or freezes the child out of fluency. The tools are part of the world these kids will work in, and a graduate who never learned to direct, verify, and distrust a model is unprepared in a different way; the goal was never zero contact but supervised, sequenced contact that builds both the internal mind and the tool fluency on top of it. Age changes the dosage: a seventeen-year-old drafting with AI and defending every choice orally is practicing the adult centaur pattern, while a seven-year-old needs almost none of it, because the early years are nearly all construction. And some children, with learning differences or without homework help at home, gain real equity from a patient explainer, provided the attempt-first spine holds. The friction is the protected resource. The tool can serve it or eat it, and the adults choose which.

Key takeaways: kids, AI, and school

The answer is conditional and the condition is friction: attempt first, AI second, explanation aloud at the end. Learning runs on desirable difficulties, retrieval, generation, productive failure, and an answer machine placed before the attempt deletes the construction while polishing the artifact. Minors are a special case because unformed faculties are not recoverable the way unpracticed ones are; so adults run the sequence, schools mediate with age gates and oral defense, and the model gets cast permanently as the questioner, never the oracle. What is being protected is the build itself, the same project, started early, as Building Your First Brain, free for the first 1,000 readers.

Frequently asked questions

Should kids use AI for school?

Yes, under one strict condition: the struggle stays. The Build First Brain rule for minors is attempt-first, AI-second: the child works the problem, writes the draft, or attempts the recall before any model is consulted, and then uses AI for feedback, questions, and correction rather than answers. Cognitive development runs on effortful retrieval and self-generated work; an answer machine deletes the very exercise that builds the mind. Used as a questioning tutor with the attempt preserved, AI helps. Used as an oracle, it quietly prevents the construction it appears to assist.

What does AI actually do to a child’s learning?

It depends entirely on where it enters the loop. Learning science is unambiguous that durable learning comes from desirable difficulties, effortful retrieval, spaced practice, generating answers before seeing solutions, and that fluent, easy input produces confident forgetting. AI placed before the attempt removes the difficulty and with it the learning; AI placed after the attempt, as feedback, explanation, and Socratic questioning, can amplify it. The tool is neutral; the sequencing is everything.

Is a child using AI different from an adult using it?

Categorically. An adult who offloads a faculty loses practice on something already built; a child who offloads it during the construction years may never build it at all. Writing, reasoning, and recall are not just skills children acquire, they are the workouts that wire the underlying circuitry. That asymmetry is why age matters in every serious policy framework, and why the friction worth removing from an adult’s workflow is often the exact friction a child’s development requires.

How should schools handle AI?

Mediate it rather than ban it or surrender to it. The direction set by UNESCO’s guidance is sensible: teacher-mediated use, age thresholds, data protection for minors, and AI literacy taught explicitly. In practice the strong classroom patterns are attempt-first sequencing, oral defense, students explain and defend work aloud, which makes outsourced thinking visible immediately, and assessment that rewards reasoning shown rather than product delivered. Bans just move use home, unsupervised; surrender moves the thinking to the machine. Mediation keeps the construction site open.

What can parents do at home?

Hold three lines. Attempt-first: homework gets a genuine try before any AI, and the child marks where they got stuck, which turns the model into targeted help rather than a ghostwriter. Explanation: whatever AI contributed, the child explains the result back in their own words, the oral test no language model can sit for them. And modeling: let them see you struggle with hard things yourself, because children learn their relationship to difficulty from the adults nearest them.