The standard complaint about students and chatbots is that the chatbot lies. It invents citations, mangles dates, and states falsehoods in the same even tone it uses for facts. The complaint is correct, and it is the smaller one. Suppose the engineers win. Suppose the error rate keeps falling until, for the kind of question a sixteen-year-old asks about a poem or a derivative or the causes of a war, the machine is reliably right. The case for keeping it out of a learner's hands does not weaken when that happens. It gets stronger. A wrong answer at least leaves a gap where doubt can grow. A correct answer, delivered instantly and in fluent prose, closes the gap and removes the work. The better the tool gets at answering, the more completely it removes the one thing the student was there to do.
That is the whole argument: the tool is most damaging when it works best. Not a defect waiting for a patch, but the design working as intended.
To see why, you have to be clear about what learning is, because the chatbot quietly assumes the wrong answer. Learning is not the transfer of a correct answer from one place into a student's notes. Decades of work in cognitive psychology, much of it associated with Robert Bjork, lands on a result people find counterintuitive and then ignore: the study conditions that make learning feel fast and smooth are usually the ones that make it stick least. Bjork's term for the conditions that work is desirable difficulties: spacing practice out, mixing problem types, and above all retrieving an answer from your own memory instead of rereading someone else's. Performance during a study session and durable learning come apart, and can move in opposite directions. The hour that feels most productive often teaches the least.
A sharper version sits directly under the argument. When you generate an answer yourself, even a slow and partly wrong one, you retain the material better than if you read a finished answer, however perfect the finished answer is. Psychologists call it the generation effect, and it has held up for decades. The reaching, the failing, and the correcting are not the inefficient path to understanding. They are what builds it, and an assistant that hands over the finished answer is built to skip it.
A study from the MIT Media Lab last year put some numbers on the intuition. Researchers had people write essays while wired to an EEG, one group using a language model, one using a search engine, one working with nothing but their own heads. Across the sessions the model group showed weaker neural connectivity, reported the lowest sense of owning what they had written, and frequently could not quote a single line from an essay they had finished minutes before. The authors called it cognitive debt. It is a small preprint, fifty-four participants on one task, and should be read as a flag rather than a verdict. But it points at something every teacher already recognises. When the machine does the composing, the composing does not happen inside the student, and whatever that act would have built does not get built.
Now put an expert and a student side by side, asking the same question and getting the same correct answer. It does opposite things to them. The expert has a structure to hang it on, and the answer drops into judgment already earned. The student is still constructing that structure, and the answer arrives in place of the act that would have built it. The same tool that amplifies the expert hollows out the learner, and not because the content differs. The harm is the substitution. This is the developmental mismatch: the people most exposed to it are the ones still building the faculties they would later use to judge the output.
The intervention is the mode, not the ban
The instinct here is to ban the technology for children, the way Australia has just banned social media for them. That is the wrong cut, because the problem is not the model but the interaction pattern: ask, receive, move on. The same underlying model, given a different instruction, runs the opposite pattern. It can take a student's question, decline to answer it, and instead probe for the exact point the reasoning breaks, returning the one hint that sends them back to fix it themselves.
This is not hypothetical. Khan Academy's Khanmigo is a GPT-class model deliberately constrained this way. Ask it to solve the problem and it will not; it asks what you think comes first and goes from there. Same technology. Opposite effect on the person using it.
So the proposal is narrow. Inside formal education, the default tool put in front of learners should be the Socratic kind, a model built to provoke the work rather than perform it, and the open answer-on-demand model should not be the thing a student reaches for to get schoolwork done. What is being restricted is a mode of interaction, not a class of technology, and that distinction carries the entire case. "Ban AI for children" throws away the best use of the technology along with the worst. "Require that AI built for learners is built to make them think" keeps it.
The part nobody has solved
That sounds clean until you try to build it, and the hard part shows up at once. Not every question deserves to be met with another question. A student who needs a date to get on with an argument, the spelling of a word, a definition they will use rather than be tested on, should simply be told. Socratising a genuine lookup is not pedagogy, it is obstruction, and nobody learns anything from being interrogated about a fact they were entitled to look up. So the tool has to triage: answer some questions plainly, push back on others, and tell which is which in real time.
That judgment is the unsolved piece, and it is worth being honest about how hard it is. A good human tutor makes the call constantly, reading the student, the moment, and where the question sits in what they are trying to learn, and gets it wrong all the time. A model has far less to go on. It cannot see whether a question is load-bearing, the thing the lesson exists to build, or peripheral, scaffolding the student has to clear to reach the real work. "What is the derivative of x squared" is a throwaway for one student and the entire point of the afternoon for another, and the difference lives in a curriculum the model cannot see and a head it cannot read.
The two ways of getting it wrong are not symmetric, and the asymmetry is the most useful thing here. Err toward answering and you quietly hand over something the student needed to struggle with; the failure is invisible, the lesson skipped without a trace, the debt building unseen. Err toward questioning and you patronise a learner who wanted a fact and got a seminar; that failure is loud, immediate, and resented. The invisible error does more damage, because nobody catches it. The visible error is more dangerous to the policy: the patronised learner closes the tab and opens the general model that just answers. Bad triage does not only fail the student. It manufactures the circumvention the policy then has to fight.
Which forces the question of who draws the line. It is a pedagogical and value judgment, not a technical one, and it belongs with the people who own the curriculum rather than being baked silently into a vendor's model where no student or teacher can see it or argue with it. A boundary set inside a company's training process is a company deciding what thinking a minor is permitted to outsource, made invisibly. The defensible version of this tool is one where the line between answer-this and work-this-out is set by the course and visible to everyone using it. That is also the hardest version to ship, which is most of why it will not get built by default.
The objections, which do not go away
None of this survives contact with the three objections that matter, and I do not think any of them can be fully answered.
The first is enforcement. Gates leak, and the freshest evidence is sobering. Australia removed roughly 4.7 million under-16 accounts in the first month of its social media ban, which took effect in December 2025. Within three months, surveys found more than half of under-16s still on the platforms, around two-thirds saying the services had taken no action against their existing accounts, and teenagers trading tips on parents' IDs and printed face masks to beat the age checks. The regulator reported no clear drop in the harm the ban targeted: complaints about cyberbullying and image-based abuse did not fall. A gate that removes millions of accounts without moving the harm indicator is doing something other than what it advertises. If an education gate only redirects the compliant while the determined route around it, what has it bought?
The second is worse, because it inverts the equity the policy is meant to serve. The learner who can get around a restriction is the one with the spare device, the VPN, the parent willing to lend an ID, the household where someone can explain the workaround. The learner who cannot is the one with none of that, and often with no human tutor either. Push hard on access and you take the answer machine away from the child who had nothing else, while the advantaged child keeps the workaround and the private tutor on top of it. A policy sold as protection for the vulnerable can remove their floor and leave the ceiling where it stood.
The third is that the boundary is porous by construction. A sanctioned Socratic tool is one browser tab from a general model that will just answer. You can only make the boundary hold by enforcing it beneath the application, at the device or the network, and the instant you do that you are deciding at the level of infrastructure which kinds of thinking a minor is allowed to reach. That is a heavy power, and it does not obviously belong to a school district, a ministry, or a vendor. The more airtight you make the wall, the worse the question of who holds the key.
When it is worth doing, and when to stop
So is it worth doing? Conditionally, and the conditions are strict enough that they double as the lines where you scope it back.
It is worth doing where you genuinely control the environment and stay honest about how short that reach is. The defensible version is not a population-wide ban on minors touching general models, which fails the way the Australian ban is failing. It is a change to the default inside formal education: the school-issued, lesson-embedded assistant is the Socratic one, and the open oracle is not what the institution hands a student to do their work. That moves the median behaviour of the compliant majority, the only thing a leaky gate can do, and it is honest only if you say as much rather than pretend it contains anyone determined to get out.
It is worth doing only if the Socratic tool is good enough that learners do not flee it. A tool that patronises loses to the one that answers, every time, and deserves to. The triage has to be good: the plain answers plain and quick, the pushback reserved for the questions that are the lesson. Get that wrong and the tool drives learners straight to the open model, defeating itself.
And it is worth doing only as provision, not subtraction. The reason to put a Socratic tutor in front of the disadvantaged learner is that they may have nothing else: no tutor, no help at home, no one to sit with them while they get it wrong. Give them a good tutor and you add a floor. Take the answer machine away and supply nothing and you remove one. The equity case is entirely contingent on the first version, and collapses into its own worst objection in the second.
The place to stop is where the only way to hold the mode boundary is to wall off the general internet for minors at the device or network level. That cure is worse than the disease: it hands someone the authority to license acceptable cognition for a generation, and no one should be granted that power quietly. A porous boundary honestly described beats an airtight one that needs a gatekeeper nobody should trust with the key.
I am not certain this nets out positive. The enforcement objection is strong, the equity objection stronger, and the cleanest version of the policy is the least likely to get built, because the Socratic tool is the hard product and the oracle is where the easy money sits. What I am sure of is the thing underneath. The harm in a classroom full of chatbots is not the mistake, and a more accurate model does not fix it; it deepens it. If we are going to hand this technology to people in the middle of building the faculties they would need to use it well, the least we can do is aim it at the work rather than at doing the work for them. The version that asks is harder to build and easier to walk past. It is also the only one that leaves the learner with more than they walked in with.