The Window of Formation

---

If you've spent any time around the conversation about AI and kids, you've probably seen the imagery. A brain draining into a phone. A skull with a funnel where the top of the head should be. Headlines that use words like "shrink" and "drain" and "lose" — the vocabulary of damage, the same vocabulary you'd use for something that already existed and got hurt. It's dramatic. It's memorable. And it's aimed at the wrong mechanism.

Damage is the wrong model here, and the difference matters more than it might first appear. Something that gets damaged was there, and now it's diminished. A muscle that atrophies from disuse was built once and is now weaker than it used to be. That's a real phenomenon, and it happens to adults who let a skill go unused for long enough. But it isn't what's most at stake for a child who grows up with AI doing large parts of the cognitive work before that child has ever done it independently.

What's at stake for that child isn't damage. It's absence. A capacity that was never built in the first place isn't the same as one that got hurt. Nothing was there to lose. The problem isn't that something got worse. It's that something never arrived.

This distinction isn't just semantic housekeeping. It changes how a worried parent should understand the situation, and more importantly, what they should actually do. Damage implies something has already gone wrong, and the natural response to something already wrong is alarm, maybe even a sense that the harm can't be undone. Absence implies something hasn't happened yet, and the natural response to something that hasn't happened yet is different: you can still make it happen. One framing puts a parent in a defensive crouch, waiting to see how bad the damage turns out to be. The other framing puts a parent in an active position, with something concrete to do.

Consider what this looks like in an actual household. A ten-year-old who has never had to sit with a confusing math word problem, because an AI tool will restate it clearly and walk through the steps the moment she asks, hasn't had anything damaged. Nothing in her has been worn down. She simply hasn't yet had the experience of sitting with confusion long enough to work through it herself, because the tool has been quietly available to remove that experience before it could happen. That's not an injury. That's an opportunity that keeps getting skipped, homework assignment by homework assignment, without anyone deciding to skip it on purpose.

This reframing isn't meant to talk anyone out of being concerned. The underlying worry that led to those tabloid images in the first place is not wrong. Something real is at stake in how children build the capacity to think through hard problems on their own. But aiming that concern at the wrong mechanism leads to the wrong response. A parent convinced their child's brain is being damaged might reach for prohibition, keeping AI away entirely, treating it as something to be minimized or hidden from. A parent who understands the actual issue as absence, not injury, reaches for something more useful: making sure the child still gets the experience of struggling with a problem before the tool steps in to resolve it.

That's a small shift in words with a large shift in what follows from them. The rest of this page will spend most of its time on the evidence for exactly this mechanism — why an unused capacity fails to form, what that looks like in a real study, and what a parent can actually do about it in an ordinary household on an ordinary evening. But it begins here, with getting the metaphor right. The real risk was never a brain being fried. It's a skill that never got the chance to form in the first place, and every day that chance isn't taken is a day it can still be.  

Before this page goes any further, it's worth being just as clear about what it isn't arguing as what it is. A title like "The Window of Formation" can suggest a stronger claim than this page actually makes if a reader isn't told plainly where the claim stops. So here's exactly where it stops.

This page is not arguing that there's a fixed age at which the ability to build a particular skill closes forever, the way a door might shut and lock. That's a real thing in some parts of biology — certain aspects of vision, for instance, depend on specific experience arriving during a narrow developmental stretch, or the underlying wiring doesn't form the same way later. But that kind of hard, biologically fixed cutoff is a different and much stronger claim than anything this page is making about thinking, writing, or professional judgment. When researchers have gone looking for that kind of fixed window in complex human capacities, what they generally find instead is more forgiving than a slammed door. Adults who never learned to read can still gain real reading skills and show measurable changes in how their brains process language after just a few months of instruction. Older adults can meaningfully improve memory skills through structured training, with benefits that last for years. None of that looks like a door that's already closed. This page isn't claiming children have a ticking clock that runs out at a specific birthday.

This page is also not arguing that AI is inherently harmful or that any use of it puts a child's thinking at risk. Most of what a person does with AI in a given day probably has nothing to do with the mechanism this page is describing. Asking it what time a store closes doesn't touch any capacity a person is in the middle of building. The concern here is narrower and more specific than "AI use is dangerous." It's about what happens when AI performs the particular thinking a person hasn't yet learned to do for themselves, not about AI as a category of technology.

This page is also not claiming that professional judgment, writing ability, or emotional regulation have been proven to depend on some special developmental window, the way a plant depends on a season. That would be an overstatement, and it's worth saying so directly rather than letting the title imply more certainty than the evidence supports. When this question has been examined in the research literature and tested through sustained, adversarial questioning of an AI system, the honest answer has consistently been the same: there's no controlled study showing that professional judgment has to form during some particular life stage or not at all. What does hold up is something narrower and better supported — that assistance which helps a beginner can become useless or even counterproductive once someone has already built real expertise in that specific area, a pattern researchers call the expertise-reversal effect, documented across dozens of studies going back decades. That's a real and well-established finding. It is not the same as proving a fixed formation window exists for judgment, or for writing, or for the ability to manage a difficult emotion.

The word "window" itself deserves a direct explanation, since it's doing a lot of work in the title of this whole page. It's being used the way a builder might use it, to describe a period during which a particular kind of practice does more to build a capacity than the same practice would do at another time. It is not being used the way a biologist uses it to describe a hard developmental cutoff, and it's not meant to suggest that once a certain age has passed, a skill can never be built at all. If that stronger meaning were what the evidence actually supported, this page would say so plainly, the same way it says plainly what the evidence does support elsewhere. It doesn't, so it won't.

Spelling all of this out here, before the rest of the page makes its case, is deliberate rather than defensive. A title carries an implicit promise about how strong the argument behind it is going to be, and a reader deserves to know upfront exactly how far that promise extends and where it stops. The claim this page is prepared to defend is specific: that timing relative to whether a skill has already formed changes what assistance does to a learner, and that this matters enough to think about carefully. That claim does not require a hard biological deadline, a blanket judgment against AI, or proof that every human capacity works this way. It only requires what the following sections are about to lay out plainly, one piece of evidence at a time.  

The developers in the Shen and Tamkin study weren't struggling with something exotic. They were doing what almost anyone does when they're learning something new and something goes wrong: staring at an error, trying to figure out what went wrong, trying something else, getting it wrong again, and eventually understanding why. That process feels like friction in the moment. It's also, as it turns out, the actual mechanism by which the skill gets built. The group that used AI encountered that friction a third as often. The group that worked without it hit the wall, again and again, and had to climb over it themselves.

This isn't a strange or technical phenomenon limited to software developers. It's the same thing that happens anywhere a person is learning to do something they can't yet do reliably.

Think about learning to parallel park. The first several attempts are usually bad. You misjudge the angle, you cut the wheel too early, you end up too far from the curb or too close to the car behind you, and you have to pull out and try again. It's uncomfortable, and it would be easy to have someone else just park the car for you every time until you eventually got your license anyway. But if someone parked it for you  single time you were learning, you wouldn't actually learn to parallel park. You'd learn to be a passenger while someone else does it. The discomfort of getting it wrong, sitting there re-angling the wheel and trying again, is not a flaw in the learning process. It is the learning process.

Or think about a new employee learning a company's internal software for the first time. The employee who gets stuck, has to click around, tries the wrong menu, eventually finds the right one, and remembers where it was next time, is building something; the employee who has a coworker walk them through every step is not building. Both of them finish the task. Only one of them will remember how to do it alone next week.

Cooking without a recipe works the same way. The first few times someone tries to season a dish by taste instead of by following exact measurements, they'll probably get it wrong; too salty, not enough acid, something a little off. That's not a failure. That's the specific experience that eventually produces the ability to taste something and know what it needs. A person who always follows someone else's exact instructions can produce a good dish every time, but they're not building the judgment that lets them improvise later, when there's no recipe to follow at all.

This is why the Shen and Tamkin result matters beyond the specific case of software developers learning a Python library. The pattern it captured—less struggle in the moment, less skill afterward—may extend well beyond coding. It shows up anywhere a task can be completed successfully without the person doing it actually having to develop the ability behind it. A student can turn in a well-written essay that AI helped structure without ever developing the ability to structure an argument alone. A new manager can send a well-handled difficult email that AI helped draft without ever building the judgment to handle the next hard conversation without help. In every one of these cases, the finished product looks the same from the outside whether or not the underlying skill actually got built. That is exactly what makes this hard to notice while it's happening, and exactly why unaided performance eventually has to be checked. 

None of this means struggle is good for its own sake, or that a person should refuse help out of some sense that suffering builds character. Nobody benefits from being handed an impossible problem with no path through it, and there's a real, well-documented pattern in learning research, called the expertise-reversal effect, showing that the exact same kind of help that's essential for a beginner can actually get in the way once someone already has real skill in an area. Help isn't the problem here. The question this page keeps returning to is narrower than that: whether the specific struggle that builds a specific skill is still happening, or whether it's being quietly and consistently removed before the person ever gets the chance to work through it themselves.

That's what happened, measurably, to half the people in the Shen and Tamkin study. It's also likely happening in ordinary households and workplaces every day, in situations that will never show up in a research paper, simply because nobody's checking.  

Eight sections in, it's worth stopping and saying the whole thing plainly, in case any of it got lost in the studies and the statistics and the parking-lot analogies.

It is not AI versus no AI.

It is whether AI performs the thinking a person needs to practice, or leaves that thinking for the person to do.

That's the entire argument. Everything before this section was building the case for it. Everything after this section will apply it, first to what "leaving the thinking for the person to do" actually looks like in practice, then to what it means for a professional, a teenager, and a child, and then to what a parent or a workplace can actually do about it. But the sentence itself doesn't get more complicated than that, and it's worth being suspicious of any version of this argument that tries to make it more complicated than it needs to be.

Later in this page, this same idea gets stated a different way, as a question of timing rather than operation, because timing is often the easier version to notice in the moment, before a task is even finished. But it's the same underlying claim, described from two angles. AI performing the thinking before someone has learned to do it, and assistance arriving before the capacity it substitutes for has formed, are two ways of describing the same event. This section states the version that's easiest to test against evidence. The later version states the one that's easier to remember at the kitchen table.

This is why the question from Part Two, whether AI use is good or bad, has never had a stable answer. It was never a question with one answer, because it was really two different situations wearing the same description. A person using AI to check work they already know how to do is in a completely different situation from a person using AI to produce work they don't yet know how to do. Both of them can be described, accurately, as "using AI." The word doesn't distinguish between them, even though the two situations produce opposite outcomes.

The Shen and Tamkin study is the clearest evidence for this because it isolated the variable directly. Everyone in the study had access to the same AI tool. The people who used it to ask questions and check their own understanding came out of the task about as capable as the people who never touched it. The people who let it handle the parts of the task they didn't already know how to do came out of it measurably less capable, even though, from the outside, both groups finished the same assignment successfully. Access to AI wasn't the variable that mattered. What the AI was being allowed to do, on the person's behalf, was.

This sentence is short on purpose because it needs to survive being repeated by someone who only half-remembers where they heard it. A parent explaining this to a spouse at dinner, a teacher explaining it to a colleague in the hallway, a manager explaining it to someone on their team, none of them are going to recite a study's statistics from memory. What they might remember, if it's stated clearly enough the first time, is the actual distinction: not whether the tool was used, but what it was used for.

It's worth being honest about what this sentence doesn't do. It doesn't tell you exactly where the line is in every situation, because that line moves depending on what the specific person already knows how to do. The exact same request to an AI tool, "help me write this," can sit on either side of that line depending entirely on whether the person asking has already built the skill of writing that particular kind of thing. This sentence also doesn't resolve the harder, still-open question the later sections of this page take on directly, whether the timing of when a skill was skipped changes anything beyond the immediate result. That's a real, separate question, and it doesn't get answered by restating this one more clearly.

What this sentence does do is give you something to check your own decisions against, in the moment, before a task is finished and it's too late to tell the difference from the outside. Before asking AI for help with something, it's worth asking one question first: do I already know how to do this, or am I about to let AI do it for me before I've ever learned how? That's not a rule. It's just the honest version of the question underneath everything else on this page.  

There's a part of this argument that doesn't depend on anything discussed so far. It doesn't need the Shen and Tamkin study. It doesn't need scaffolding, substitution, or any claim about timing at all. It only needs one observation about how junior jobs have always worked, and what happens when that arrangement quietly breaks.

For as long as entry-level work has existed, it has produced two things at once, even though only one of them ever showed up on an invoice. The first is the actual work: the drafted contract, the debugged code, the researched brief, the audited spreadsheet. The second is something nobody had to pay for separately, because it came bundled in automatically: a person who now knows how to do that work a little better than they did before they started. Every tedious, junior-level task a young associate, analyst, or engineer grinds through wasn't just labor being extracted from them. It was also, whether anyone planned it that way or not, an apprenticeship. Call the first product Output A. Call the second Output B.

Nobody had to manage Output B on purpose. It arrived as a natural side effect of a junior person doing junior work under some supervision, making mistakes, getting corrected, and slowly becoming the senior person who trains the next junior person after them. This is, more or less, how every profession has replenished its own expertise for as long as professions have existed. The system didn't need a formal training department to produce Output B. Producing Output A reliably produced Output B as a byproduct.

Generative AI changes this arrangement in a very specific way, and it's worth being precise about exactly what changes. AI can now produce something very close to Output A directly. It can draft the contract, generate the code, and assemble the research brief at a fraction of the time a junior employee would need. What it cannot do is hand over Output B along with it. There is no version of an AI system that leaves behind a slightly more experienced human being as a byproduct of having done the work. The organization can now get the finished product. It does not automatically get the person who would have been formed by producing it.

This is not, on its own, an argument that AI is bad for business. From a narrow, immediate view, it can look like a straightforward improvement. The work gets done faster, often at lower cost, without the errors a genuine beginner would have made along the way. Any single decision to use AI instead of assigning a task to a junior employee can be perfectly rational in isolation, exactly the same kind of individually reasonable decision described back in Part Four's account of what happens across a whole institution making similar rational choices at once. Nobody has to make a mistake, or act in bad faith, or ignore an obvious cost, for this pattern to take hold. It happens through ordinary, sound decision-making, repeated enough times.

The cost is what happens next, and it doesn't show up on the same balance sheet as the savings. Output A gets captured immediately. Someone bills for the finished contract, the deployed code, and the completed brief. Output B's absence isn't a cost anyone has to record anywhere, because accounting has never had a line item for the expert who didn't get formed. That cost surfaces later, sometimes five years later, sometimes ten, when the organization needs someone who has actually done the junior-level version of a task hundreds of times, developed real judgment about how it can go wrong, and can now supervise, mentor, or catch a mistake nobody else noticed. If fewer people spent those years doing the work directly, there are fewer people who can do that.

This is worth stating in the plainest terms possible: the person who decides to use AI instead of a junior hire today is very often not the person who will need that missing expert ten years from now. A law firm partner who lets AI draft routine contracts instead of assigning them to first-year associates is not necessarily the one who will suffer when there's a shortage of seventh-year associates capable of spotting a subtle liability clause. A hospital administrator who streamlines diagnostic support with AI today is not necessarily the one in the room when a rare case needs a doctor who built real pattern recognition the slow way. The decision and the consequence can be separated by enough time and enough organizational distance that nobody making the choice today ever has to see what it costs.

None of this requires believing anything about brains, formation windows, or the timing of learning. It only requires three things most people would accept without much argument: expertise is built through accumulated, hands-on experience; AI can now perform a meaningful share of the tasks that experience used to come from; and there's no guarantee any organization will deliberately replace the training that used to happen automatically as a side effect of getting the work done. Put those three together, and a real problem exists whether or not the rest of this page's argument about timing and formation windows turns out to be right. That's not a weakness in the case. It's the reason this argument can stand on its own, even if every more speculative claim on this page were eventually disproven.

Part Eleven ended with a claim, stated in a single paragraph: the apprenticeship-unbundling argument doesn't need anything else on this page to be true. It's worth slowing down and actually testing that claim rather than just asserting it again, because if it holds up under real pressure, it changes how much weight the rest of this page has to carry.

Imagine, for a moment, that every developmental claim on this page turns out to be wrong. Imagine the order-effect experiment described later gets run, and it finds nothing, that assistance arriving before or after a skill has formed makes no lasting difference once total practice is held constant. Imagine Shen and Tamkin's finding fails to replicate, or turns out to be specific to coding and nothing else. Imagine, in other words, the strongest version of the null hypothesis wins completely, and it turns out timing, in the specific sense this page has been arguing, doesn't matter at all.

Does the apprenticeship argument survive that?

It does, and it's worth being precise about why. The apprenticeship argument was never a claim about how skills form inside an individual brain. It never depended on a critical period, a sensitive period, or any biological mechanism at all. It rested on three much plainer observations, none of which have anything to do with developmental timing. Expertise gets built through accumulated, hands-on experience, doing a task repeatedly, under some supervision, over years. AI can now perform a meaningful share of the tasks that experience used to come from. And no organization is required, by any law of economics or any law of nature, to deliberately replace the training that used to happen automatically as a side effect of getting work done. Every one of those three claims stands independently of whether "the window of formation" turns out to be a real phenomenon, an overstated one, or no phenomenon at all.

This is actually a well-established pattern in how economists think about human capital, long before AI entered the picture. The economist Gary Becker's foundational distinction, between skills specific to one employer and general skills useful across an entire field, was built decades ago to explain something that has nothing to do with cognitive science: why firms underinvest in training that a competitor could later benefit from by simply hiring the trained employee away. That's a market-structure problem, not a brain-development problem. It would be just as real in a world where human cognition worked completely differently than it does, because it isn't about how a mind learns. It's about who captures the value of the learning and who pays for it.

That's the real test of whether an argument is load-bearing on its own. A claim depends on another claim if disproving the second one would force you to abandon the first. Try disproving the formation-window hypothesis completely, and see what happens to the apprenticeship argument. Nothing happens to it. The three observations it rests on don't move. A junior associate who never drafts a hundred routine contracts still won't develop the pattern recognition that comes from drafting a hundred routine contracts, whether or not there was ever a specific window during which that pattern recognition needed to form. The mechanism doesn't require a window. It only requires that experience be the thing expertise is made of, which is a much harder claim to argue against than anything about developmental timing.

This matters for a practical reason beyond intellectual tidiness. The formation-window hypothesis is genuinely uncertain, and this page has said so plainly, more than once. A reader who isn't persuaded by it, who thinks the evidence is too thin, the sample sizes too small, the order-effect question too far from being settled, doesn't have to reject everything else on this page along with it. The apprenticeship argument doesn't ask for that kind of trust. It asks for something much smaller: that experience builds expertise, that AI is beginning to displace a real share of that experience, and that nobody has guaranteed a replacement for what's being displaced. Anyone who accepts those three plain claims has already accepted the argument, whatever they eventually decide about windows, timing, or anything happening inside a single learner's mind.

That's why this section exists separately from the rest of the page, rather than folded quietly into a footnote. Not because it's the most dramatic claim here. Because it's the argument most likely to still be standing in ten years, regardless of what the research on formation windows eventually shows.

Part Fourteen ended by naming the experiment nobody has run. This section does something different, and it's important to be precise about what: it doesn't guess at what that experiment would find. Guessing would be exactly the move this page has spent fourteen sections refusing to make. What this section does instead is lay out the three ways the timing hypothesis could turn out to be true or false, and what each one would actually mean, without picking a favorite.

Outcome one: it disappears easily

It's possible that timing, in the strict sense Fourteen defined, turns out not to matter much at all. In this outcome, someone who got AI assistance before mastering a skill, and someone who got the identical assistance after, end up in roughly the same place, once both have put in the same total amount of independent practice. What looked like a lasting deficit in Shen and Tamkin's study would turn out to be something closer to what a lender would call training debt: a real shortfall, but one that gets paid down quickly once genuine practice resumes, not a mark that stays on the record permanently. If this is what a real order-effect study eventually found, the practical response would be modest. Notice the gap, add some deliberate unaided practice, and move on. Nothing about how AI gets used would need to change dramatically. The scaffolding-and-substitution finding would still matter, since substitution would still cost a person real practice time in the moment, but the specific fear this page's title points toward, that something is being lost, permanently, by getting help too early, would turn out to be smaller than it looks.

Outcome two: it's recoverable, but expensive

A second possibility sits in between. In this outcome, the deficit is real, and it doesn't erase itself through ordinary practice, but it can still be closed with enough deliberate, targeted effort, more than it would have taken to build the skill correctly the first time. Think of it less like a debt that pays itself down and more like a debt that requires a specific, harder repayment plan than the original loan would have. Someone who used AI to skip debugging early on might eventually become a competent debugger, but it might take a structured remediation effort, mentorship, or deliberately AI-free practice sessions to get there, where building the skill the first time around might not have required that same extra push. If this is the true outcome, the practical response is bigger than the outcome one's. It means noticing gaps matters, but it also means organizations, schools, and individuals need an actual plan for closing them, not just an assumption that time and continued work will do it automatically.

Outcome three: it doesn't close

The third possibility is the one that would make the word "window" in this page's title fully earned. In this outcome, even giving both people the same amount of later practice doesn't close the gap. The order in which the experience arrived would have done something that later effort can't fully undo. This is the strongest version of the timing hypothesis, and it's also the one with the least direct evidence behind it right now. If it turned out to be true, the implications would be serious enough to justify real caution now, before the evidence exists to confirm or rule it out, because by the time it's confirmed, an entire generation of learners could already be past the point where the finding would help them.

Why laying this out matters, even without an answer

None of this is a prediction. Nothing in this section says which of these three is more likely than the others, and it shouldn't be read as leaning toward the most dramatic one just because it's the most attention-grabbing. What this section is actually doing is something narrower and more useful: showing that the stakes of these three outcomes are not the same size. Outcome one asks for a shrug and a little unaided practice. Outcome two asks for a real, structured response. Outcome three asks for something closer to the caution this page has been building toward from the start. When the possible outcomes of an open question are this different in size, it becomes reasonable to act with more caution than the current evidence alone would justify, not because the worst outcome is confirmed, but because the cost of being wrong about it is so much higher than the cost of being wrong about the others. That asymmetry, not a guess about which outcome is true, is the actual argument for taking this seriously now, while the experiment that would settle it still hasn't been run.

Fifteen showed why an unresolved question can still carry real stakes. Sixteen and Seventeen showed what it looks like to watch for evidence responsibly instead of waiting for certainty that may never fully arrive. This section is where all of that turns into something usable: an actual answer to the question a reader has probably been holding since Part One. Given everything this page has laid out, what am I supposed to do?

It's worth pausing on a word this page has used loosely until now: proof. Nothing on this page has been sitting in the dark waiting for evidence to show up. There's real evidence throughout, a tested finding in Part Six, decades of research on expertise and assistance in Part Ten, and three independent datasets pointing in the same direction in Part Seventeen. What's missing isn't evidence. It's the one decisive experiment, described precisely in Part Fourteen, that would turn the timing hypothesis specifically from a well-supported idea into a settled fact. Confusing those two things, treating "no decisive experiment" as "no evidence at all," would misrepresent everything the last several sections have just built.

That distinction resolves something that might otherwise look like a contradiction. If the standard for acting were proof, and proof only arrives after an experiment has already run its course, then nothing could ever justify caution while it still mattered, since proof, by its nature, shows up after the window for acting on it has closed. That's not how caution actually works anywhere else. A fire alarm doesn't wait for proof of a fire. It responds to evidence, smoke, and heat, sufficient to make acting the reasonable choice, weighed against how cheap the response is and how costly it would be to ignore. This section applies the same ordinary logic here: not proof as the bar for action, but evidence, growing and checkable, weighed honestly against what it would cost to be wrong in either direction.

The honest answer isn't a single rule. It depends on a distinction that shows up constantly in fields that have to make decisions under real uncertainty, such as medicine, engineering, aviation, and financial regulation, and it's worth borrowing that logic directly rather than inventing something new. The decision that makes sense depends on two questions.

How easily could this risk be corrected later, if it turns out to be real?
How expensive is it to guard against the risk now, before knowing for certain?

Everything else follows from those two questions.

When the risk is reversible and easy to check, proceed and watch. If a mistake would be quickly visible and simple to fix, the sensible move is to go ahead and pay attention, not to freeze in place waiting for proof. A parent letting a teenager use AI to check their own already-completed math homework, work the teenager did independently first, falls here. If it turns out the checking habit is somehow interfering with something, that would show up fast, in grades or in a simple unaided test, and it could be corrected immediately. Demanding a randomized trial before allowing this would be a strange, disproportionate response to a small, easily monitored risk.

When the risk is harder to reverse but cheap to guard against, use the safeguard. Some situations don't resolve themselves as quickly, but protecting against them doesn't cost much either. This is where most of the practical advice on this page actually lives. A household rule like showing your own attempt before consulting AI costs almost nothing to follow and directly preserves the one thing this page has argued matters most, the chance to struggle with something before help arrives. When a safeguard is this cheap relative to the risk it's guarding against, the responsible move is to just use it, rather than waiting to find out for certain whether the risk was real.

When the risk is systemic, slow to reveal itself, and hard to undo, ask for more before removing the last safeguard. This is the category that matters most for the harder, less personal decisions this page has discussed, the apprenticeship argument from Part Eleven especially. If an entire profession quietly stops training its own junior members, and the deficit doesn't become visible for five or ten years, correcting it after the fact isn't like flipping a switch back. It means rebuilding a training pipeline that took decades to establish in the first place, for professionals nobody can simply hire away from somewhere else, because nobody was training them anywhere. When a potential mistake looks like that, slow to reveal itself and expensive or impossible to reverse once it does, the reasonable standard shifts. It's no longer "wait for proof." It becomes "don't eliminate the last safeguard until there's a real reason to believe it's safe to."

It's worth being precise about what this framework does not say. It doesn't say AI should be restricted, regulated, or avoided as a general matter. Most of what happens when a person uses AI falls into the first category: low stakes, easily checked, correctable if something's wrong. It doesn't say every organization needs a formal training-preservation policy starting tomorrow. It says something narrower and more useful: that the size of the caution should match the size and reversibility of what's actually at risk, not the volume of the conversation around it. A great deal of the public discussion about AI and thinking treats every use of AI as equally risky, which is exactly the mistake this framework is built to avoid. Checking already-completed homework and eliminating an entire profession's junior training pipeline are not the same category of decision, and they shouldn't be treated as if they were.

This is also where the three outcomes from Part Fifteen and the framework from this section connect directly, worth stating once plainly rather than leaving it implicit. Fifteen asked what each outcome would mean if it turned out to be true. This section asks how much caution each outcome would justify. Put together, they answer the same question from two directions. If the timing hypothesis turns out to be the mild, easily-closed version, outcome one, the cost of having taken modest precautions now will turn out to have been small. If it turns out to be outcome three, a durable, hard-to-reverse effect, the cost of not having taken those same precautions could be significant and, by the time it's confirmed, largely unfixable. That asymmetry is the entire argument for erring toward caution on the third category specifically, not because outcome three is the most likely of the three, but because it's the one where waiting for certainty costs the most if the caution turns out to have been warranted after all.

The next section takes this framework one step further. If this page is going to ask for real attention now, before the evidence is complete, it's worth being specific and public about exactly what would eventually prove it wrong. 

Everything on this page up to now has been about understanding the problem precisely. This section is about the part that's actually usable at a kitchen table or in a classroom on an ordinary Tuesday. Not a warning. Not a rule about limiting AI. Something closer to a habit, built around the one thing Part Ten already established matters most: whether a person gets to do the thinking, or whether it gets done for them before they've had the chance.

The single most useful thing a parent can do costs nothing and takes no extra time. Ask to see the attempt before the AI gets involved. Not as a punishment, and not as suspicion, just as the default order of operations. A kid working on a math problem tries it first, on paper, even badly, before opening a chat window. A teenager drafting an essay writes a rough first paragraph in their own words before asking AI to help shape it. The point isn't the quality of that first attempt. Most first attempts are messy, and that's fine. The point is that the attempt happened at all, because that's the moment this page has repeatedly argued deserves protecting. A household rule this simple, show your own work first, protects the one thing this entire page has been arguing matters, without requiring anyone to become an AI skeptic or track screen time.

Schools can do a version of the same thing, at a slightly larger scale. The instinct many schools have reached for is detection, trying to catch AI-written work after the fact, which puts a school in an adversarial position with its own students and treats every use of AI as cheating rather than distinguishing scaffolding from substitution the way Part Ten does. A more useful instinct is redesigning what gets assessed in the first place. An in-class, unaided piece of writing, done by hand or on a locked-down device, tells a teacher something a take-home essay increasingly can't: whether the student can actually produce the reasoning themselves, not just recognize good reasoning when AI hands it to them. This isn't about banning AI from schools. Plenty of AI use in a classroom is genuine scaffolding, help understanding a concept, a second opinion on an argument's weak point, exactly the kind of use Part Ten already said has a real, legitimate place. The goal is making sure some meaningful portion of a student's work still happens the way Shen and Tamkin's control group's work happened, unaided, with real errors, real correction, and real practice at recovering from being wrong.

There's a second thing worth doing that's less about rules and more about attention. Every so often, without warning and without making it feel like a test, ask a kid or a student to do something similar to what they just did with AI's help, except this time without it. Not as a gotcha. As information. If the unaided version comes reasonably close to the AI-assisted version, that's a good sign, evidence that the underlying skill is actually there and the AI is functioning the way Part Ten describes scaffolding functioning, supporting a capacity that already exists. If the unaided version falls apart completely, that's worth knowing now, while there's still time to do something about it, rather than years from now, the way Part Eleven's apprenticeship argument describes professions finding out too late.

None of this requires becoming an expert in AI, tracking every interaction, or treating every homework session like a research study. It requires one shift in instinct: noticing that a finished piece of work, on its own, doesn't tell you whether a skill got built or got skipped, the same blind spot Part Three named at the very start of this page. The finished essay looks the same either way. The only way to tell the difference is to occasionally look at what happens when the help isn't there.

That's a small, repeatable habit, not a solution to everything this page has raised. But it's the one piece of this entire argument that doesn't require waiting for 2031, doesn't require a randomized study, and doesn't require resolving whether the timing hypothesis turns out to be true. It just requires making sure the struggle still happens somewhere, regularly, on purpose, before the tool that could remove it gets the chance to.

Nobody is going to get this exactly right every time, and this page was never asking for perfect. It was asking for one thing.

Notice who's doing the thinking.

Everything else follows from that.

Stay Sovereign.

— Jim Germer  

August 11, 2026