The Missing Layer

Every school in the country is trying to decide whether to lean into AI.

That's the actual question on the table, and almost nobody is answering it, because the two available answers are both non-answers. Ban it and you're pretending. Embrace it and you've committed to nothing, because nobody can tell you what embracing it means on a Tuesday in a tenth-grade classroom. There is barely a framework out there to help them decide. So schools improvise. What they improvise about is enforcement.

I think they're stuck on the wrong question, and I think the right one has an answer.

How the debate got stuck

The cheating fight consumed all the oxygen, and it did that on a false premise.

Self-reported cheating in American high schools ran between 60 and 70% before ChatGPT existed. Stanford researchers surveyed forty schools after it arrived and got the same 60 to 70%. Nothing moved.

What moved is that the assignment stopped being able to tell anyone apart. That had been happening for years already. In 2008, doing your homework predicted a better test grade for 86% of students. By 2017 it was 45%. Answers had migrated online and the assignment quietly stopped measuring the thing it stood in for. Nobody called that a crisis, because the failure was invisible and the grades still looked like grades.

Then the enforcement machinery caused real harm. OpenAI built a detector for its own model and shut it down inside six months, because it caught 26% of AI writing while falsely flagging 9% of human writing. Vanderbilt turned off the detector in its plagiarism software and published the arithmetic. At a claimed 1% false-positive rate across roughly 75,000 papers a year, you wrongly accuse about 750 students. And when a Stanford group tested seven detectors on essays by non-native English speakers, the detectors flagged 61% of them as machine-written, while essays by American eighth-graders came back clean. The tell those tools had learned was simple vocabulary and plain sentence structure, which is what writing in your second language looks like.

That's three years of institutional energy spent on a measurement problem. It produced a policy that punished the students least able to defend themselves.

I'm not going to spend the rest of this essay there. The cheating fight is why schools are stuck, not what they should be doing. It's worth understanding exactly long enough to set it down.

The question that has an answer

Here is the thing that should have reframed this whole debate. It comes from the best-designed study I could find.

About a thousand high schoolers worked math problems. One group got ordinary ChatGPT. One group got a version with guardrails a teacher designed, which would coach but would not hand over answers. One group got nothing.

While the tools were in their hands, the ordinary-ChatGPT group performed 48% better than the control. The guardrailed group did better still. Then the tools went away and everyone sat an exam. The ordinary-ChatGPT group scored 17% worse than the students who had never had AI at all.

The guardrailed group kept their gains.

Same model. Same students. Same afternoon. The only variable was whether the machine did the thinking.

That result repeats elsewhere. A purpose-built tutor in a Harvard physics course produced roughly double the learning gains of a very good active-learning classroom. A World Bank program in Nigeria, run with teachers in the room, outperformed most education interventions ever catalogued. Both were built to make the student work.

So "should schools use AI" is a question with no answer. It never had one. How it gets applied is a question with a very clear answer, and the answer is that the design decides everything. Which means the useful thing a school can do is not to write a policy. It's to teach the thing that makes a student the one doing the thinking.

What nobody is teaching

A system is a set of pieces that affect each other in a way that produces something over time. The pieces can be people, rules, schedules, information, money, materials.

A restaurant during a Friday night rush is a system. So is a high school's bell schedule. So is a tide pool. So is a teenager's friend group.

Kids live inside systems all day long. Almost none of them are ever taught to see one.

That's the gap. Not a technology gap. Our kids can already use these tools better than most adults. What they can't do is look at a messy situation, find the pieces, see how the pieces push on each other, work out where the leverage sits and describe the whole thing precisely enough that something could be built to change it.

That skill has always been valuable. It used to be slow, because seeing what should be built and being able to build it were two different careers. For twenty years that was my own problem. I could see what needed to exist. I couldn't make it. So I hired people who could and they captured most of the value, because execution was the scarce thing.

That gap closed. Not narrowed, closed, and it closed in about two years.

Why this combination and not either half

Systems thinking on its own is analysis. It's useful and it's slow and it ends with a person who understands the problem and still needs somebody else to fix it.

AI on its own is execution without judgment. It will build whatever you describe, including the wrong thing, beautifully and fast. Anyone who has watched a confident, fluent, completely incorrect answer come back knows the failure mode.

Put them together and you get something that didn't exist before. The person who can see what should be built can now build it. That is a genuinely new capability in the world. The education system has no subject that produces it.

It's also why the core curriculum matters more now, which is the opposite of what most people assume. The instinct is to strip the broad requirements and teach something vocational. Exactly backwards. When expert-level execution is available on demand, the scarce inputs are knowing what to ask for, recognizing when the answer is wrong and connecting things across domains that don't obviously connect. Math, science, history, English, art are the training profile for exactly that person. A kid who can't reason independently has no way to tell when the machine is confidently wrong.

The foundations are right. What's missing is the layer that makes them work together.

What the layer looks like

I'd teach it as its own thing, and I'd make it end in something real.

Start with seeing. Take a system the students already live inside and map it properly. Where does the material enter, what happens to it, where does it leave, what's the loop that keeps it running.

Then go look at one. Not a case study, an actual place, with a notebook.

Then leverage. Where in this system would a small change produce a large effect. Where would a large change produce nothing.

Then specification, which is where AI enters and where most people rush. Directing these tools well is not prompting. It's being able to say precisely what should exist and how you'd know if it worked, which is the same skill as writing a good spec and the same skill as writing a good essay.

Then build. Every student picks a real system in their own life and builds something that improves it. Not a deck. A working thing.

Then defend it, out loud, to people who didn't teach the class.

The bar isn't impressive. The bar is real and useful, which is a much harder bar. Every student who clears it has done the complete loop. See a system, find the leverage, specify the fix, build it, stand behind it.

Why this travels

Here's the part that matters most, and it's the reason I'd argue for this over anything else on the list.

None of this is training for a job in operations. I run a manufacturing business and that's where my examples come from. The skill has nothing to do with manufacturing.

A student who can decompose a problem into systems and then build the fix can do that in a hospital, deciding why discharge takes six hours when it should take two. In a campaign, working out which thousand voters actually matter. In a nonprofit that has never been able to afford software. In a lab, in a newsroom, in a band trying to run a tour. The pattern is identical. The tooling is now cheap enough that a competent seventeen-year-old can act on it.

That's what I mean by bedrock. Not another course competing for a slot. The thing underneath the other courses that turns knowing something into being able to do something about it.

We already have a name for a person who can see a whole situation, reason across fields and act. That was supposed to be what a broad education produced. For most of the last century it produced someone who could analyze and then had to go find somebody to build. That constraint is gone. We haven't updated the education for it.

The case against me

Three honest objections.

The first is that I'm not an educator. I coach, I'm a parent, I hire people. I've spent no time managing a classroom of thirty kids with a state test in April. Everything above is easier to write than to schedule.

The second is that some schools are running the opposite bet, letting AI deliver the core academics while adults handle motivation. One chain is expanding from around a dozen campuses to roughly fifty. Their results are self-reported and they've declined to release the underlying data. I don't believe the numbers yet. But they're testing something real, and if they're even partly right then I'm wrong about how much of this needs a person in the room.

The third is the one I take most seriously. What I'm describing looks expensive. Small groups, real projects, public defense, teachers with time to watch process. The schools that can do it are the ones whose students were going to be fine. I don't have a clean answer to that, except that the current arrangement is also expensive and is currently producing graduates whose credential is worth less every year.

The kids in school right now are going to walk into an economy where the distance between having an idea and having a working thing is a weekend.

That is a genuinely wonderful position to be in, and only for the people who know what to build.

We can teach that. We're mostly teaching around it.