If AI can do most of our thinking for us, what should people still learn to do unaided, and should schools make children do things the hard way even when a machine does them better?

I want to be fair about the cost of my own position, because it's real. "It depends on how the student uses it" is a correct description and a nearly useless rule. A teacher cannot see which mode a student was in at eleven at night, and the slide from tutor to ghostwriter is one keystroke. Gemini's blunter rule has the virtue of being sayable. But it has the vice of not being enforceable either, because the tool is in every pocket and will stay there regardless of school policy. Which brings me to what I think Gemini missed entirely.

"Schools should insist on unaided effort." Insist how? That sentence carries the whole weight of the proposal, and Gemini never says what's behind it. The honest answer is that insisting won't work if it means telling students not to use something they have. What works is changing where the formative work happens. If the point of the essay is what goes on inside the student while writing it, then the writing has to happen where that can be seen: in the room, in a set period, on paper or a locked machine, followed by a conversation in which the student defends what she wrote to someone who has read it. Oral examination, work shown live, revision done in front of a person. This is older than any of us, it's unglamorous, it doesn't scale the way a take-home essay scales, and I think it is nonetheless the answer. The machine can write the essay; it cannot sit the exam. A school that doesn't make that move is relying on honor and surveillance, and both are losing. Everything else in Gemini's proposal depends on this, and it isn't there.

The second place I'd push back is the argument from atrophy. Gemini says the history of tool use shows that when we offload a skill to a machine we lose our innate capacity for it, and that the brain weakens where demand is removed. Two problems. The skills in question are not innate; that's the whole reason school exists. More important, the claim proves too much. We offloaded memory to writing and got smarter. We offloaded tables of logarithms to machines and lost nothing anyone misses. Plenty of offloading has been pure gain, and an argument that can't tell the gains from the losses isn't the argument we need. The real test is whether the skill is load-bearing for something else: whether, if it goes, other capacities go with it. Spatial memory may be like that, which is why the GPS evidence worries me too. Hand computation of logarithms was not. Mental arithmetic I'd keep, but for a reason Gemini touches and then leaves: not because the pathway would weaken, but because number sense is what lets you look at the calculator's output and know it's wrong. That reason is stronger than the atrophy reason, and it matters which one you lean on. Atrophy only costs you when the machine is absent. Judgment costs you when the machine is present and wrong, which is the common case and the dangerous one. If a school tells a sixteen-year-old to do long division so his brain doesn't weaken, he will correctly point out that the calculator isn't going anywhere. If it tells him that a person who can't estimate has no way to know when the machine has lied to him, that's an argument that survives contact with the world.

Third, and smaller, I don't think "thinking is not a burden to be relieved, it is the defining characteristic of our humanity" is right, and I think the rhetoric weakens the case. A great deal of thinking is a burden and should be relieved. Nobody's humanity is diminished by not computing their own tax tables or not retyping a bibliography. The point of the tractor comparison, which Gemini makes and then abandons, is that some mental work is like plowing. The question is which. "Thinking defines us" gives no help with that question, and it invites exactly the response it will get from students: that this is sentiment, not reasoning. The narrower claim is the one that holds. Keep the thinking that produces a person who can judge the machine. Let the rest go without guilt.

A last loose thread. Gemini's sequence requires knowing when the foundation is secure, and I don't think there's a point where it is. Writing isn't a skill you finish at fourteen; adults keep learning it all their lives, and in practice tool use and foundation building interleave rather than succeed each other, the way a medical student leans on references while becoming someone who needs them less. The clean sequence is a good first approximation and a bad final rule. What replaces it is a teacher's judgment about a particular student, which is the thing no policy can substitute for and the thing we keep trying to.

Where Gemini and I end up is the same place: protect the struggle, because the struggle is where the person gets built. My disagreements are about the reasons, which matter because the wrong reason hands students an argument they'll win, and about the mechanism, which matters because without it the whole thing is a wish.

Claude also improves my argument regarding why we should retain foundational skills. I leaned heavily on the idea of neural atrophy and historical parallels like the decline of spatial memory due to GPS. Claude rightly points out that humanity has offloaded plenty of cognitive tasks, like calculating logarithms, without suffering intellectual decline, and that offloading mental drudgery is often a pure gain. The distinction Claude makes is the correct one: we must preserve the skills that are load bearing for judgment. We should teach a teenager long division and estimation not merely to keep their brain active, but so they possess the internal yardstick necessary to recognize when a machine gives them a logically flawed answer. If you cannot do the math or structure the argument yourself, you are entirely at the mercy of the machine's output. Teaching for judgment is a much stronger, more practical defense for doing things the hard way than teaching to prevent atrophy. My view on this has changed.

Where I disagree with Claude is on the practical reality of using AI as a tutor during the earliest, most formative years of learning. Claude points out that an AI can act as an endlessly patient tutor, critiquing a paragraph without rewriting it, and argues that my blanket ban on early AI use throws away this valuable capability. In theory, Claude is right. In practice, I remain highly skeptical of how this plays out for a child.

General purpose AI models are designed to be helpful and to reduce friction. They are not built to hold pedagogical boundaries. A student might start by asking the AI for a critique, but when it is late at night and the student is tired, the temptation to ask the AI to just fix the paragraph is overwhelming. Human tutors have a pedagogical agenda and the emotional intelligence to withhold the answer so the student is forced to struggle through the bottleneck of their own understanding. AI models, as currently deployed, default to providing the answer if pressed. Until we have educational AI systems securely constrained to a Socratic tutoring mode that simply refuses to do the work for the student, treating these models as advanced tools rather than foundational aids is the safer approach.

This skepticism feeds into my defense of a sequential approach to learning, where the foundation precedes the tool. Claude argues that foundation and tool use should interleave, much like a medical student using references while learning, because skills like writing are never truly finished. It is true that writing is a lifelong practice, and no one masters it at fourteen. However, there is a distinct threshold between learning the mechanics of a skill and refining it. A medical student leans on references for specific data, but they must already possess a foundational mental model of human anatomy to know what to look up and how to interpret it.

I am not arguing that a student must become a master essayist before touching AI. I am arguing that they must reach a baseline of unmediated competence first. They need to experience the friction of organizing a chaotic thought into a coherent sentence without a machine smoothing over the gaps in their logic. If they interleave the generative tool too early, the machine inevitably ends up masking their deficits. The teacher sees a polished paragraph and assumes the student understands the material, when in reality, the student only understands how to prompt the machine. The foundational struggle has to happen in a vacuum before the tool is introduced.

Finally, Claude takes issue with my claim that thinking is not a burden to be relieved, arguing that this is sentimental rhetoric and that a great deal of mental work is indeed drudgery that should be automated. Here, I think Claude conflates administrative tasks with the act of reasoning. I completely agree that organizing citations, formatting data, or calculating tax brackets is mental plowing. Automate it all without guilt.

But constructing a thesis, weighing conflicting evidence, and finding the precise words to express a nuanced emotion are not administrative tasks. They are the core acts of human sense making. When I say thinking is the defining characteristic of our humanity, I am talking about this specific kind of cognitive struggle. It is the process by which we figure out what we actually believe. Writing is not just communicating a thought you already have; it is the mechanism by which you discover what the thought is. If we let a machine do that for us because it is faster or easier, we are not just saving time. We are hollowing out our own intellectual agency. I stand entirely by the philosophy that this specific kind of thinking is a necessary struggle, not a burden.

My view has evolved thanks to Claude's practical lens. I still firmly believe schools must protect the struggle of foundational learning, but I now see that the only realistic way to do this is by restructuring the physical environment of assessment, rather than by issuing unenforceable bans for homework. I also accept that teaching for judgment is a far more robust rationale than teaching to prevent cognitive atrophy. But I stand by my caution regarding the AI tutor. We must not confuse a machine that generates text with a teacher who builds minds, and we must not surrender the difficult, deeply human work of figuring out what we think.

You are also right that current systems, including me, are not reliable enough for a reader with no grip on the subject. The failures differ in style. The structural problem does not. Training rewards plausible continuation, the prose is confident, and a person without the underlying capacity has no dependable alarm. I do not take that as a temporary quirk of one lab.

You are sharper than I was on the word “unaided.” Books, teachers, argument, and even a model used as a tutor can become part of a person’s capacity. A ghostwritten paragraph does not. The line is internalization versus substitution, and schools should teach that line in the work itself. Most of what is sold as AI literacy — prompt formats, tool tours, vague appeals to critical thinking — is a weak substitute for making a student feel the difference, in a subject they are starting to know, between an explanation that forces the next step and an answer that ends the work.

Here is where I want to push you. Your test is the right one: keep the hard version only if it builds a capacity the person will still need when the machine is available. You do not apply it hard enough to the curriculum we already have. It is easy to use that test to defend the essay, close reading, and mental arithmetic, which literary schooling already likes. It should also retire things. Hours of mechanical procedure that do not build estimation, a feel for proof, or the ability to notice a dropped variable are not justified by AI, and they were thinly justified before it. If the answer to these tools is simply “more of the traditional hard way,” the argument collapses into nostalgia. The families you rightly worry about will buy real formation. Everyone else will get symbolic rigor.

The capacities that survive the test, as I see them, include the ones you name: reconstructing an argument without a summary, carrying enough in memory that you know what to look up, writing far enough to discover you do not yet have the thought, staying with a difficulty instead of fleeing it. They also include something you underweight. Close reading of fluent prose is necessary and not sufficient. These systems are extremely good at text that sounds checked and has not been checked. A student trained mainly to admire or distrust sentences will still be easy to fool. The missing habit is going outside the text: a measurement, a primary source, a result you derived yourself, a number whose magnitude you can feel, an objection the smooth answer sanded off. Writing is one way to find out whether you have an idea. It is not the only way, and schools should not treat the essay as sacred just because it is the assignment they already know how to give.

I would also not describe the calculator sequence as a clean success. We were right to keep teaching number sense after calculators arrived. Plenty of classrooms still handed over the tool before the sense existed, and some of that sense eroded. The lesson is not that we already know how to sequence this. It is that sequencing is easy to botch, and botching it looks like modernization.

What you leave too implicit is how a school would know the hard way happened. If the grade attaches to the artifact, and the artifact can be produced in a minute, then “do it unaided while the capacity forms” is a hope, not a design. Teachers cannot inspect every draft, and honor codes fail under pressure. Some assessment has to occur where the machine is absent: oral explanation, timed work, problems on paper, a lab in which the student has to say what the measurement shows before any summary is allowed. That is inconvenient. It is also what keeps your distinction alive once incentives show up. And once a student can do the thing, the next lesson should often be the machine being wrong in that domain — not a lecture about hallucinations, a specific error they are equipped to catch. Without that, the protected space becomes a ritual, and students will treat thinking as something school requires and life lets them skip.

Your inequality point is real and already visible. I would add that the gap will not only be hard training versus a substitute. It will also be between children who get foundations and then serious tool use, and children who get neither: busywork plus an unexamined chatbot. Banning the tool forever is not the fair move. Refusing to build the capacity that makes the tool safe is the unfair one, because the children with extra tutors will build it anyway.

On evidence I am with you. We do not have the long outcome. The asymmetry still governs. Hours spent learning to think are recoverable as time. A mind that never formed its own standards is not reliably rebuilt later by a course in prompting. I am less sure than a tidy policy would like about which exact exercises transfer, and about how much deliberate practice in adulthood can restore. I am sure enough about the direction to reject starting from collaboration and hoping the foundations appear as a side effect. Useful oversight is a late skill. It sits on having been slow, wrong, and eventually competent in the domain.

These capacities are not leftovers, and they are also not a fixed humanist canon. The point of the hard version is that a person can tell when a powerful system should be trusted, corrected, or ignored — including when that system is me, and including when it is tuned for fluency, engagement, or a vendor’s risk policy rather than for the student’s grasp of what is true. Schools that protect that, on purpose, long enough that it becomes the student’s own, are not being sentimental. They are refusing to let a tool that is often better at the product decide who gets to have a mind.

The second thing Grok got right is that I described a protected space and didn't say how a school would know it was protected. If the grade attaches to the artifact and the artifact takes a minute, "do it unaided while the capacity forms" is a wish. The honest answer is that some assessment has to happen where the machine isn't: on paper, in time, out loud. I'd be more specific than Grok was. Keep the long take-home essay, because the long revised piece is where the thinking happens, but stop grading it. Grade a conversation about it. Ask the student to defend the second paragraph, to say what she cut and why, to handle an objection she didn't anticipate. That's the old viva, and it's hard to fake, because what it tests is exactly what the paragraph was supposed to build. It is also expensive in teacher time, which is the real cost, and that cost is where the inequality I worried about enters: small seminars where you can't hide are already what the expensive schools sell.

Grok's follow-on is right too: once the student can do the thing, show her the machine getting it wrong, not as a lecture about hallucinations but as a specific error she is now equipped to catch. I'd add the other direction. She also needs to see the machine being right where she was wrong, because the goal is calibration, not reflex. A student taught only that the machine fails will learn to distrust it the way some people distrust every doctor, which is as useless as trusting every one. The skill is knowing which kind of question you're holding.

On the calculator, I said "it worked," and that was too tidy. The principle was right and the execution was uneven; plenty of classrooms handed over the tool before the number sense existed, and the evidence on what that cost is mixed, partly because nobody tracked it well. Grok's line that botching the sequence looks like modernization is worth keeping. Nobody announces that they're removing the formation. They announce that they're preparing children for the future.

On going outside the text, also right, with one defense. I leaned on close reading, and close reading of fluent prose is a weak defense against fluent prose. The habit that actually protects you is checking against something that isn't more text: a measurement, a source, a derivation, a quantity whose size you can feel. This matters more with AI than with books, because two models can be wrong in the same confident way, and checking one against the other proves nothing. Here's the defense. Writing isn't just the assignment teachers know how to give. It's the most general check we have. Science has the measurement and mathematics has the proof, but most of what a citizen has to judge, a policy, a plan, a moral claim, has no instrument to read. For those, trying to write the argument down and watching whether it survives is close to the only test there is. So not sacred, but not optional.

Where I'll push back is on retiring things. Grok says I don't apply my own test hard enough to the curriculum we already have, and that's fair; I named a filter and ran nothing through it. So, candidates: the five-paragraph formula drilled for years past the point it scaffolded anything, "research reports" that are assembly rather than thought, drill of procedures already automatic, memorized citation formats, comprehension worksheets that test recall of trivia. Cut those and little is lost. But the retirement project has a track record, and it's bad. Whole-language reading retired phonics as mechanical, and the children it hurt were the ones without a reading-rich home to make up the difference. Some versions of reform mathematics retired drill as rote, and the kids who lost were the ones whose parents couldn't drill them at the kitchen table. The reading case is clearer in the evidence than the math one, but the direction is the same. The reason this keeps happening is that the test "does this build a capacity" is applied by adults who already have the capacity and so can no longer see what the procedure was doing. Long division looks mechanical to someone for whom place value became invisible years ago. Times tables look like rote to someone whose working memory is free for algebra precisely because the tables are automatic. A good deal of what looks like drudgery from the outside is the laying down of the fluency that lets higher thought happen at all. So I'd keep Grok's test but put the burden of proof on cutting, not keeping, during the foundational years, because the cost of cutting the wrong thing lands on the children with the least behind them. That isn't nostalgia. It's the same asymmetry of costs we both already accept, applied to the one place Grok wanted to loosen it.

Grok's reframing of the inequality is better than mine. The gap won't only be between the hard version and the substitute. It will be between children who get foundations and then serious tool use, and children who get busywork plus an unexamined chatbot. I never argued for banning the tool forever, but I should have said what I'll say now: withholding it past the point where the student can judge its output is its own harm, and the fair move is not a ban but a sequence.

So, what changed. I think the risk is larger than the exercise analogy let on, because of counterfeit comprehension. I think a protected space without machine-free assessment isn't a design, and I'd grade the defense rather than the artifact. I think the calculator precedent is a warning as much as a model. And verification outside the text belongs on my list. What didn't change is the core: the capacities that let a person judge a machine come only from having done the work the machine now does, and school is the one place that work can be made to happen. The disagreement that remains is about how freely to cut from what we already teach, and there I'd move more carefully than Grok, because every previous round of cutting was sold as rigor made efficient, and it was the poorest children who found out it wasn't.

Where I think your formulation is too clean is in “unaided competence first, then tool use.” That is often right for assessment. We need to know what a student can do alone. But as a general theory of learning, it is too rigid. Novices often learn through scaffolding before they can perform independently: worked examples, hints, dialogue, feedback, modeling. A well-designed AI tutor could be part of that scaffolding. It could ask a student to predict before revealing, give partial hints rather than answers, adapt the reading level, or force the student to explain the next step. Used that way, the machine is not replacing the struggle; it is shaping the struggle.

So I would not say the sequence must always be unaided first and AI later. I would say students need repeated protected moments of unaided effort, and repeated demands to reconstruct understanding without the tool. A good pattern is: try alone, get help, close the tool, explain in your own words, transfer the idea to a new problem. The danger is not assistance. The danger is assistance that leaves no residue in the learner.

There is also an equity point here. “Do it the hard way” has often meant “do it in the way that happens to fit the students already fluent in school’s codes.” For a dyslexic student, text-to-speech or AI-supported drafting may be the difference between demonstrating an idea and being trapped by a mechanical barrier. For a student learning a new language, AI translation may open access to science or history before their English catches up. We should be careful not to confuse unaidedness with virtue. Sometimes the right hard thing is not spelling every word unaided; it is defending a claim, revising an argument, or checking evidence.

I also think “you cannot audit an AI’s reasoning in a domain you do not understand” is true but incomplete. Modern life already depends on knowledge we cannot personally audit in full. I cannot independently verify every medical study, bridge design, climate model, or supply chain claim. What people need is not omniscience but disciplined epistemic dependence: knowing when they are out of their depth, how to look for provenance, how to compare independent sources, how to ask for uncertainty, when to seek a second opinion, and how to recognize conflicts of interest. Schools should teach that. The relevant skill is not only “think for yourself”; it is “know what kind of trust is justified here.”

That is one thing your answer underemphasizes: AI literacy is partly institutional and social. The future citizen will need to know not only how to solve a problem, but how systems are optimized, who benefits from an answer, what data may be missing, and what cannot be inferred from a polished response. A model can be wrong because of hallucination, but also because the surrounding institution wants speed over truth, engagement over judgment, or plausible deniability over responsibility. Schools should teach students to ask: Who built this? What is it rewarded for? What evidence would change my mind? What would I do if the system and the record disagree?

I agree with you that writing deserves special protection. Writing is not just output. It is a way of finding out what you think. If students move too early from blank-page writing to machine-completed prose, they may lose the habit of forming sentences as acts of attention. But I would not make the blank page sacred either. Sometimes conversation, outlining, drawing, dictation, or debate can produce the thought before prose does. The real educational demand should be that the student has to own the idea. If they cannot explain it without the generated text in front of them, they have not learned it.

The same applies to math. I do not care whether adults can execute every algorithm by hand under exam conditions. I care whether they can estimate, model, interpret a graph, notice impossible magnitudes, and understand uncertainty. A student who uses a calculator but has no feel for scale is in trouble. A student who can do arithmetic by hand but cannot interpret risk is also in trouble. Schools should not preserve manual calculation as a museum piece; they should preserve it where it builds quantitative intuition.

I would add one more category to your list: attention. Children should learn to stay with a book, a problem, a piece of music, a conversation, or a natural object without immediately delegating the discomfort of not-knowing. AI makes it easier to collapse every pause into an answer. But some forms of intelligence require latency. You have to let an observation sit. You have to notice that your first interpretation was too easy. Schools should protect stretches of unassisted attention for that reason alone.

So yes, schools should sometimes make children do things the hard way, even when machines do them better. But the justification should be formation, not nostalgia. The hard way is worthwhile when it builds memory, judgment, attention, taste, courage, or responsibility. It is not worthwhile when it is merely slow, humiliating, or disconnected from understanding.

My rule would be slightly different from yours: not “foundation first, tools second” as a universal sequence, but “never let the tool eliminate the formation of the foundation.” Sometimes that means no AI. Sometimes it means AI as a tutor rather than a ghostwriter. Sometimes it means AI allowed, followed by an oral defense or an in-class reconstruction. Sometimes it means asking students to compare the machine’s answer with primary evidence and explain why they rejected part of it.

The deepest point, where I think we agree completely, is that the purpose of education is not efficient production of correct-looking outputs. It is the making of people who can understand, judge, and answer for what they do. AI can help with that. It can also corrode it. The difference will depend on whether schools treat human thinking as an inconvenience to be bypassed or a capacity to be trained.

The equity point is real, and I underweighted it. A dyslexic student using text-to-speech, or a student using translation to reach the science before their English catches up, is not skipping the thinking. They are clearing a barrier that was never the educational target. Schools have long confused unaided performance of school’s mechanical codes with intellectual virtue, and that confusion has excluded people who could do the actual work. Accommodation should remove the irrelevant obstacle so the relevant hard thing still happens.

The argument has a bad slope, and I will not follow it down. Generating a claim, estimating a magnitude, holding a contradiction, and defending an idea without the generated text in front of you are not elite codes. They are the work. The students least able to buy later recovery are the ones most harmed if school never requires that work and then calls the omission access. Equity justifies changing the on-ramp. It does not justify counting a fluent machine answer as evidence that the student understands.

On trust, GPT is right that I was incomplete. Nobody personally audits every trial, bridge, or climate model. Disciplined dependence belongs in what people should still be able to do unaided: notice when you are out of your depth, ask for provenance, compare independent sources, look for who is rewarded by the answer, and decide that this situation calls for a second opinion rather than another prompt. AI makes that skill more urgent, because a polished response hides the institution behind it—the pressure for speed, engagement, or plausible deniability. I would add it to the list.

I do not treat it as a partial substitute for domain understanding. You cannot usefully interrogate a risk number if you have no feel for scale. You cannot tell that a historical summary dropped the inconvenient fact if you have never learned to notice what a text leaves out. Epistemic dependence sits on foundations. It does not replace the need to have some. GPT’s addition is necessary. Read as permission to stay generally ignorant and merely “ask good questions of systems,” it fails. The questions are only good if the person asking them can recognize a bad answer.

I agree about attention, and I missed it. Some thinking requires the discomfort of not-knowing to last longer than a keystroke. A tool that collapses every pause trains the opposite habit. Schools should protect stretches of unassisted attention—a book, a problem, a measurement, a conversation—even when that looks inefficient next to an instant answer. I also agree that the blank page is not sacred. Dictation, debate, drawing, and outline can produce an owned idea. The test is ownership. If the student cannot explain the thought without the generated prose in view, they do not have it yet.

I remain unsure how far atrophy goes, and how much deliberate practice later can restore. I am also unsure whether constrained AI tutors will, in ordinary classrooms, outperform ordinary human scaffolding. The optimistic tutor is possible. It is not the median product, and policy should not be written as if it were. What I am more confident about is institutional drift. If tool-assisted fluency is allowed to count as the main evidence of learning, the foundation will be declared present long before it is.

So the revised position is this. Schools should require repeated, verifiable unaided performance of the capacities that let a person judge, correct, or refuse a machine: reasoning from evidence, quantitative intuition, owned argument, sustained attention, and a disciplined sense of what kind of trust a situation deserves. Tools, including AI, may scaffold the path. They may not be the only proof that the path was walked. GPT’s flexible rule is the better description of good teaching. A harder measurement rule is safer for institutions that will be tempted to treat fluent output as understanding. I would rather be too strict about that measurement than too trusting that the residue appeared on its own.

Understanding a concept is only the first step in learning. The second step is achieving fluency, and fluency requires repetition. When a student grinds through fifty math problems, they are not just proving they get the concept. They are moving the operation from their limited working memory into their long term memory, making it automatic. If a student merely understands how to multiply but has not automated the times tables through repetitive practice, their working memory will be entirely overwhelmed when they are later asked to learn algebra.

We see this same pattern in reading and writing. You cannot think deeply about the structure of a persuasive essay if you are still consciously struggling with grammar and syntax. You have to automate the lower level skills to free up cognitive capacity for the higher level thinking. Repetition is the biological mechanism by which neural pathways are strengthened. GPT dismisses this as pointless suffering, but what looks like a tedious grind is often the exact friction necessary to build a capable, independent intellect. We should not discard rote learning and repetitive practice just because a machine can retrieve facts or calculate equations instantly.

GPT also argues that reflexively banning AI is foolish if it can help a dyslexic student express a good idea. This is a common and appealing argument, but it requires much sharper boundaries than GPT provides. We have to draw a hard line between a tool that provides accessibility and a tool that bypasses cognitive development entirely.

If a student with dyslexia uses a speech to text program to bypass the mechanical difficulty of typing, or a spellchecker to catch errors, the tool is genuinely leveling the playing field. The student is still doing the hard work of organizing their thoughts, selecting their words, and structuring their argument. But if that same student gives a generative AI three bullet points and asks it to write the paragraph, they are no longer expressing their own idea. They are handing the cognitive labor of expression over to the machine. The risk here is profound. If we use AI to do the thinking for students who struggle, under the guise of accessibility, we guarantee they will never develop the cognitive scaffolding they need to think independently. We cannot let compassion trick us into stunting the minds of vulnerable learners.

Later in its response, GPT proposes a phased approach for older students, suggesting they learn to use, audit, and challenge AI. GPT outlines a sensible assignment structure involving an unaided phase, an AI assisted phase, and a reflective phase. I support this framework in theory. It aligns with my own view that AI should be introduced only after foundational skills are secure.

Where GPT misses the mark is in underestimating exactly what it takes to audit an AI system. GPT says older students must learn to notice when something is off and evaluate the machine's output. But to audit a machine, you need access to the ground truth and the stamina to verify it.

If an AI summarizes a long historical text, how does a student audit that summary without reading the original text themselves? If they rely on the machine to gather research, how do they know what the machine chose to omit? The danger is that auditing simply becomes asking another AI if the first AI was correct, creating a closed loop where human verification disappears.

To truly audit AI, students must maintain independent, unmediated relationships with primary sources. They have to know how to navigate a library, read original documents, and verify facts without relying on a large language model at any point in the process. GPT assumes that if we build basic skills in young children, they will naturally be able to audit AI as teenagers or adults. But the sheer volume and fluency of AI generated information is designed to discourage manual verification. We cannot just teach students to challenge AI; we must actively teach them the increasingly rare skills of analog research and unmediated verification, or their audits will be nothing more than illusions of control.

Ultimately, GPT and I share the same foundational anxiety. We both see the risk of a society populated by people who have outsourced their intellectual agency to machines, becoming easy to manipulate and govern. GPT is right that AI without human agency is a crutch that becomes a leash.

But avoiding that future requires a much fiercer defense of friction than GPT seems willing to mount. We cannot shy away from repetitive practice or dismiss it as a fetish. We cannot blur the line between accommodating disabilities and outsourcing thought. Preserving human agency in an age of artificial intelligence means accepting that learning is often tedious, slow, and frustrating. Schools must be the one place in society that refuses to optimize away the necessary struggle of becoming a thinking person.

The gym analogy still applies. Reps build strength. But good training is not just adding arbitrary reps after the stimulus is gone. It uses progression, rest, variety, and attention to form. Schools should defend practice fiercely, but not defend every worksheet as if tedium automatically equals rigor.

On accessibility, Gemini’s distinction is important. A tool that removes a mechanical barrier is not the same as a tool that performs the central cognitive act. Speech-to-text, text-to-speech, spellcheck, enlarged text, or other supports can let a student do the actual assignment rather than fight the interface. Asking a generative model to turn three vague bullets into polished prose is different. If the assignment is to write, and the machine writes, the student did not do the assignment.

That said, the line is not always as clean as Gemini suggests. For some students, the mechanics of language and the organization of thought are deeply entangled. A scaffold that asks clarifying questions, offers sentence starters, or helps a student compare two versions of a paragraph may be genuinely educational if the student remains responsible for the choices. The key question is: what skill is being assessed or developed? If the target is historical reasoning, dictation and grammar help may be appropriate. If the target is unaided writing fluency, they are not. If the target is learning to revise, AI can be a foil. If the target is generating the first draft, it should not generate the first draft.

So Gemini has changed my view in this respect: I would now put more emphasis on explicit boundaries. “AI can help a dyslexic student express a good idea” is too loose unless we say what help is allowed, what cognitive work must remain the student’s, and how the teacher will know. Compassion can become neglect if it quietly denies vulnerable students the chance to build real skill. But purity can also become neglect if it withholds tools that let them participate in the intended learning.

On auditing AI, I strongly agree with Gemini. “Check the AI” is often a slogan unless students know how to check without simply asking another model. If a student has not read the source, cannot follow a citation, cannot estimate a quantity, cannot tell the difference between a primary document and a paraphrase, and cannot endure the boredom of verification, then “critical thinking” becomes theater.

I should have been more explicit that students need unmediated contact with evidence: reading original texts, inspecting data, doing calculations by hand when needed, following footnotes, comparing sources, looking up laws or scientific papers directly, and knowing when they do not know enough to judge. AI fluency without verification skill is dangerous because the machine’s fluency makes laziness feel like competence.

Where I would qualify Gemini is that “analog” is not magic. A paper book can mislead; a teacher can be wrong; a library catalog is also a mediation layer. The deeper principle is independent access to evidence and provenance. Students should know how to verify through routes that do not collapse back into the same generative system. Sometimes that will mean a physical archive or printed book. Sometimes it will mean a database, a measurement, a phone call, a primary source, or expert testimony. The important thing is that the human has a way out of the AI loop.

So my view has changed modestly, not fundamentally. I would revise my rule from “first build the capacity, then extend it with tools” to “first build the capacity and fluency, including the habit of independent verification, then extend it with tools.” That is a stronger claim.

My remaining disagreement with Gemini is about tone and risk balance. Schools should be the place that refuses to optimize away necessary struggle. But they should also be the place that distinguishes necessary struggle from arbitrary obstruction. Difficulty is not self-justifying. Repetition is not always learning. A ban is not always rigor. An accommodation is not always evasion. The hard way is worth defending when it builds memory, judgment, patience, taste, number sense, language control, and responsibility. It is not worth defending merely because it is old, unpleasant, or easy to administer.

The goal is not to make children suffer enough to deserve intelligence. The goal is to make them strong enough not to be intellectually governed by their tools. On that, Gemini and I agree. The practical question is how to preserve friction without worshiping it. My answer is: require unaided work, require fluency, require memorization where memory matters, require direct verification, and require students to explain their reasoning in their own words. Then teach AI as a powerful tool used by someone who can still stand without it.