By Austin VornhagenPicture a classroom.
Don’t try hard. Just let one appear. Now: where’s the clock?
Above the board, probably. Front and center, where everyone facing forward can see it. The desks face one way. There’s one adult. Something, a bell or a buzzer or a polite chime, is about to end the period whether anyone has finished thinking or not.
You just pictured the room in the film above. So did your parents. So would a kid in 1910.
The popular explanation is that school was built as a factory: rows, bells, and obedience, to train workers for the assembly line. It’s a satisfying story. It’s also incomplete, and the incomplete part is where the interesting stuff hides.
Industrial ideas about efficiency did shape schools. But so did a new republic that needed citizens who could read, reformers who wanted poor kids to get what rich kids got, cities that suddenly had too many children for one room, colleges that couldn’t read each other’s transcripts, and, later, courts and Congress forcing open doors that had been shut on purpose.
Your kid’s school isn’t one design. It’s eight old fixes stacked in a trench coat.
Each fix solved something real. Each one left furniture behind. And in 2026, a machine that can write the essay, finish the worksheet, and produce the lab report on request is leaning hard on the one thing all eight quietly depend on.
To see what that is, we have to dig.
Every piece of classroom furniture is a fossil of a fix
Tyack and Tobin had a name for the stubborn shape of school: the “grammar of schooling.” Grades by age. Subjects in separate periods. Credit counted in hours. One teacher, one room. Like the grammar of your native language, it’s so familiar you stop noticing it, and it’s brutally hard to change, because every part leans on every other part.
So let’s excavate. Scroll down and the layers settle in the order they were laid. Each one is a problem, the fix that solved it, and what the fix left behind.
Eight layers of fixes that built the American school, oldest at the bottom
Local districts and school boards
Colonial kids learned through churches, apprenticeships, tutors, and charity schools, and access depended on family money, location, race, and gender. After independence, schooling picked up a political job: a republic needs citizens who can follow public affairs and resist tyranny. Education stayed a state and local responsibility, with federal help like land grants.
Left behindOpportunity and policy change at the district line.
Free schools, paid for by local taxes
If school depends on parents’ money, many children get very little of it. The common-school movement, associated with Horace Mann, pushed for tax-supported schools with no tuition, to teach literacy and character and to bring different social classes into the same room.
Left behindThe local tax base matters. State aid can close the gap, by amounts that vary a lot.
One curriculum, and it’s mandatory
Massachusetts passed the first statewide compulsory-attendance law in 1852, though early enforcement was uneven. Once every child must attend, somebody has to decide what every child learns.
Left behindA permanent fight over whose knowledge and values belong.
Grades sorted by age
Growing cities had far more children than one mixed-age room could handle. Graded schools grouped pupils by age and proficiency, cut the curriculum into yearly stages, and gave each teacher one class. Historians David Tyack and William Tobin describe it as a deliberate invention to make schooling efficient, replicable, and fair.
Left behindOne pace for a room full of different kids.
Credit counted in hours
High school was turning into a mass institution: between 1910 and 1940, the share of American youths enrolled in secondary school rose from about 18% to 71%. Colleges needed to read transcripts from everywhere. The Carnegie Unit, introduced in 1906, standardized a course by time: roughly 120 hours of study.
Left behindFinishing the hours can hide what wasn’t learned.
A school picked by your address
A nearby school for every child is a sensible rule. But the promise of universal education coexisted with deliberate exclusion. Brown v. Board of Education rejected legally mandated segregation in 1954, and implementation met organized resistance, including school closures and private academies set up to dodge integration.
Left behindHousing patterns walk straight into the classroom.
School becomes a safety net
Hunger, poverty, and disability can stop a child from learning at all. Title I (1965) sent extra money to schools serving many low-income children. The 1975 law that became the Individuals with Disabilities Education Act established major rights for children with disabilities. Meals, counseling, and health support expanded.
Left behindMore jobs for schools, without always more people to do them.
Tests with consequences
Enrollment and graduation numbers could hide weak learning, especially for groups that vanished inside school averages. A Nation at Risk (1983) framed it as a threat to the country. No Child Left Behind (2002) attached consequences to test results; the Every Student Succeeds Act (2015) replaced it and kept key accountability rules.
Left behindA narrow number starts steering the teaching.
A machine that can make the homework
No fix yet. A tool that can produce the essay, the worksheet, and the lab report on request isn’t one more layer on the pile. It presses on all eight at once, because all eight assume the work a student hands in tells you something about the student.
Every layer solved a real problem. Every layer left behind a new one.
Now notice the part that makes school so hard to reform: the layers are glued together.
Try changing one. Switch from yearly grades to advancing when you’ve mastered something, and you’ve also changed schedules, staffing, report cards, transfers, and graduation. Redraw a district line and you’ve changed who gets which school and who controls it. Change the test and you’ve changed how families, colleges, and governments read every school in the country.
That’s why new technology tends to change what happens inside the room without changing the room. Look back at the film. The overhead projector, the computer counter, the laptop cart: each arrived as a revolution, and each got absorbed into the grammar.
“the great equalizer of the conditions of men”
That’s not a factory manager talking. In that same 1848 report, Mann described education as a way to give people enough independence to resist exploitation, and imagined schools producing inventors, farmers, scholars, and public leaders. Economic productivity was part of the pitch. It wasn’t the whole pitch.
But two of those layers deserve a closer look, because they decide the shape of your kid’s actual Tuesday.
Same hours. Very different kids.
Grades by age and credit by the hour combine into a choice every teacher faces at the end of every unit. Some students understand it thoroughly. Some partly understand it. Some are missing something the next unit needs. The calendar says: move on.
That’s not a story about teachers who don’t care. It’s what the structure produces, even with great teachers. Here it is in a toy model.
A toy model of twenty students moving through four units on a fixed calendar, then on a mastery schedule
Everyone gets the same time
Twenty students, each learning at their own speed. The tall bars show how much of each unit a student actually masters before the bell. The thin bars on top show how much time they got: identical, by design.
The bell rings. The class moves on.
The unfinished part of Unit 1 doesn’t vanish. It becomes a missing prerequisite (the red part of the bar) and makes the next unit harder for whoever carries it.
Gaps compound
Students with gaps learn the new unit more slowly, which creates bigger gaps. The kids at the front pull further ahead. Same room, same hours, and the spread keeps widening.
The transcript says everyone took the course
Every student sat through the same hours and can leave with the same course credit. The credit records the time. The bars record the learning. They are not the same thing.
Hold learning fixed. Let time vary.
Now nobody moves on until they’ve mastered the unit. Every bar fills. But look at the time bars: some students need far longer than others. Who teaches the kid who’s done in October and the one who needs until April? That’s the bill the calendar was avoiding.
A time-based school holds time constant and lets learning vary. That trade was made on purpose.
It made schools schedulable, transcripts readable, and teachers assignable. Those are real benefits; a system serving tens of millions of children has to run on something. But it means a diploma has always been, partly, a receipt for time served.
Hold onto that word: receipt. Nobody can see learning directly, so school has always run on receipts. There have been two main kinds:
The credit, the seat time, the completed course. Proof that you were there.
The essay, the worksheet, the lab report, the test. Proof that you made something.
Remember them. One of them is about to break. But first, the layers that weren’t built for everyone.
The fixes that weren’t for everyone
The trench coat has darker seams, and the story doesn’t make sense without them.
Public education’s expansion was something excluded people fought to win. After emancipation, Black communities actively built schools and pursued education as a route to independence and equality. Yet racial segregation restricted access by law, and when Brown v. Board struck legally mandated segregation down in 1954, organized resistance followed: some places closed schools, and private academies opened to evade integration.
Native children faced another kind of coercion. Federal boarding-school policies separated children from their families and suppressed their languages and identities. The Interior Department’s own investigation documents it as an intentional assimilation policy with enduring consequences.
Ending a discriminatory rule doesn’t undo the conditions the rule helped create.
The most persistent mechanism is also the plainest. Your address.
Neighborhoods and district boundaries decide school assignment, and local property taxes help fund schools. State money can offset the differences, but how much varies a lot from state to state. The Government Accountability Office has documented how housing, wealth, district boundaries, and school resources stay tangled together.
Then, from the 1960s on, schools took on stronger duties to the children they’d neglected or excluded: Title I money, disability rights, meals, counseling. Those were answers to real barriers. My read is that they also created a lasting design problem: schools gained broader responsibilities without uniformly gaining the staff, money, and flexibility to carry them. Ask a teacher about it. Bring snacks.
In 2026, school got its own report card
Here’s where the trench coat starts to wobble. In September 2026, Gallup reported that only 32% of Americans are satisfied with U.S. K–12 education, the lowest in its 27-year trend and down 11 points in two years. A separate Gallup survey asked how schools are doing at specific jobs. So let’s grade the school the way it grades kids.
| Subject | Rated excellent or good |
|---|---|
| Adapting to new technologies | 44% |
| Preparing students for college | 32% |
| Using AI effectively | 30% |
| Preparing for the world of work | 21% |
| Critical thinking and problem solving | 20% |
Weakest marks in the subjects that matter most after graduation.
That last line is the interesting one. People think schools nationally are slipping while rating their own child’s school far better. The dissatisfaction is real. It doesn’t mean most parents think their own school is failing completely.
Perceptions aren’t proof, so check the actual work. On the 2024 national assessment of twelfth-graders, released in September 2025:
scored below NAEP Basic in reading
scored below NAEP Basic in math
Both are the highest shares recorded. “Below Basic” means below NAEP’s threshold for partial mastery. It does not mean those students can’t read or do arithmetic.
The pandemic hit hard: chronic absenteeism nearly doubled between 2018 and 2022, according to the Education Department. Recovery is uneven. The 2025 long-term NAEP results show nine-year-olds improving in reading and math compared with 2022, while thirteen-year-olds showed no significant improvement over 2023 and stayed below pre-pandemic performance.
These numbers measure a problem; they don’t prove one cause. They also undercut the idea that school’s old job is obsolete. Lots of kids still need the old job done better, even as new jobs pile on.
So who feels ready? A 2026 Gallup survey of 1,758 students aged 14 to 19 found a split by path:
Share who felt prepared for the future. Overall, 60% felt prepared, up from 43% in 2023, so this measure is improving.
Then there’s the grading disagreement. Battelle’s 2026 survey of 3,629 people asked whether today’s curricula prepare students to apply their skills in real settings:
Confident that current curricula prepare students to apply skills in real settings. Stakeholder judgments, not tests of graduates.
Same students, two very different grades. Employers emphasized continuous learning, adaptability, communication, and problem-solving. (Employers’ own training habits deserve scrutiny here too.) Meanwhile, in a nationally representative 2026 survey of public-school teachers, only 18% said they’d received formal guidance on using AI at work; 48% got informal guidance and 34% got none.
And the old promise that more school reliably buys a better future? Gallup found a record-low 31% of Americans calling college “very important,” down from 70% in 2013. Yet the latest Bureau of Labor Statistics figures still show bachelor’s-degree holders earning a median $1,578 a week in 2025 versus $966 for high-school graduates, with unemployment of 2.8% versus 4.3%. Those averages don’t prove the payoff of every degree. People are questioning the price and reliability of the route. They haven’t discovered that education stopped paying.
The gap isn’t between school and its old purpose. It’s between attending school and being ready for adult life.
AI didn’t break school. It broke the receipt.
A Census Bureau working paper using business data from late 2025 and early 2026 found that 23% of firms reported workers using AI for work tasks. The leading uses of generative AI: writing, document analysis, and information search.
Now open a homework folder. Same list.
For a century, Receipt 2 worked well enough. If a student handed in a clear essay, the student could probably write a clear essay. That inference was never perfect (parents and older siblings have “helped” for generations), but making the product took effort, and the effort left learning behind.
Now the product is cheap. A polished document is weak evidence of understanding when software can produce one in seconds. And Receipt 1, hours in a seat, never proved understanding in the first place.
Both of school’s receipts now print without any learning attached.
Here’s where intuition goes wrong. Surely a student who practices with a brilliant AI helper learns more? Researchers including Wharton’s Hamsa Bastani tested exactly that in a field experiment with nearly a thousand high-school math students in Turkey. One group practiced with a plain chatbot. One practiced with a tutor version designed with guardrails: hints, not answers. A control group had no AI. Then everyone took an exam alone.
With AI in hand, both AI groups crushed the practice problems. Now take the AI away.
Performance relative to the no-AI control group. Bastani et al., field experiment with roughly 1,000 high-school students. “About the same” means no statistically meaningful difference from control.
The chat logs explain it: most plain-chatbot students simply asked for the answer. They aced practice and then did worse alone than students who never had AI at all. The guardrailed tutor largely avoided that harm. It’s one setting and one tool, not a universal verdict, and it cuts both ways: a randomized study in a college physics course found students learned more from a carefully designed AI tutor than from the active-learning class it was compared with.
Performance with the tool is not ability without it.
The lesson isn’t “AI good” or “AI bad.” It’s that the design decides, and that any school still trusting the old receipts can no longer tell the difference.
So I’d rebuild school around a different question
Suppose you got to start over, knowing machines will keep getting more capable. What is school for?
First, the trap to avoid: building education around a list of tasks humans will always do better than AI. Nobody can write that list with confidence, and every new model would rewrite it. A purpose that depends on beating the machine comes with an expiration date.
Ask instead what a person should be able to understand, decide, and do no matter how capable the machines get. Here’s the purpose I’d give the whole system:
Help people develop the knowledge, judgment, relationships, and practical power to shape their own lives and take part in shaping their society.
A machine solving a problem doesn’t give a person the ability to understand their situation, decide what matters, or influence what happens to them. Those are the things school should build. They break into six jobs:
- 1Understanding
A working grasp of language, mathematics, nature, history, and institutions.
- 2Independence
Making decisions, knowing when to ask for help, and not becoming wholly dependent on one person, company, or system.
- 3Relationships
Cooperating, handling disagreement, caring for others, belonging to communities.
- 4Practical power
The experience of turning an intention into a result with your own abilities and the tools at hand.
- 5Participation
Understanding power, and practice influencing the decisions that affect you.
- 6A good life
Enough art, movement, ideas, nature, and ways of living to discover what you actually value.
That last job deserves as much attention as job prep. A society could become extraordinarily productive while lots of people struggle to find belonging, direction, or enjoyment. Education would still prepare people for work. It would also prepare them for a future where work changes a lot and might take up less of life.
Knowledge doesn’t get demoted
The tempting conclusion is that if AI knows everything, kids don’t need to. Learning research points the other way. The National Academies’ review of how people learn emphasizes prior knowledge: you need to know things to understand explanations, ask useful questions, and notice when something’s off. General “critical thinking” can’t carry the curriculum alone.
Someone who knows little biology struggles to evaluate a claim about vaccines, however thoughtful they are. Someone who understands percentages catches a misleading financing offer before consulting any tool. So everyone still gets a rigorous common curriculum: language, mathematics and uncertainty, science, history and institutions, technology and AI, health and everyday life, cooperation, the arts, and learning itself, with deep pathways into calculus, lab science, trades, and anything else a student wants to master. What changes is what students do with it.
The new core subject: deciding what to delegate
“Knowing how to use AI” is too small a goal. The real skill is a set of questions you ask before handing anything off:
- 1What outcome am I trying to achieve?
- 2Which parts should I do myself, because doing them builds my abilities or matters to me?
- 3Which parts can I hand off?
- 4What information and authority does the system need?
- 5What evidence would justify trusting the result?
- 6Who bears the consequences if it’s wrong?
Picture an AI that recommends a school attendance policy. Average attendance improves, and children with chronic illnesses get penalized. A smarter model might produce a better policy, but someone still has to decide what protections any policy must respect. UNESCO’s AI competency framework for students already reaches beyond tool operation into human agency, ethics, and system design. I’d go further and put those questions into decisions students actually make, all year.
An honest limit: kids can’t become experts who personally verify every conclusion of a superhuman system. So school teaches how warranted trust works: evidence, independent expertise, testing, accountability, and a way to appeal. Not the fantasy that clever prompting turns you into an auditor of everything.
The big idea: let kids help run something real
Here’s the part I think is genuinely new. Keep the strong academic foundation. Then give students increasingly real responsibility for small, bounded pieces of the place they spend their days. Organize school as a small society that children gradually learn to take part in and govern.
A younger child helps plan a class activity or tends a garden. An older student organizes an exhibition, investigates a local environmental question, or helps decide how a project budget gets spent. The school supplies the materials and adult support, so access doesn’t depend on whose parents have equipment, contacts, or a free afternoon to drive.
Every substantial project runs on six steps: set a purpose and name who’s affected; learn what you need; consider alternatives; decide what work and authority to delegate; carry out a bounded plan; examine the results and revise. Teachers still teach explicitly, with worked examples, practice, and feedback, and decide when students need instruction before they investigate. Projects don’t replace teaching. They give it somewhere to land.
Watch one run.
A student project on cafeteria waste, step by step
Cut lunch waste. Nobody goes hungry.
A group of students gets a small budget and a real cafeteria. The goal has two halves: reduce what gets thrown away, and make sure every kid still gets enough to eat.
Weigh the bins
They weigh what’s discarded, separate edible food from packaging, and log attendance and menus. Along the way: sampling, measurement error, nutrition, cost, and how to interview someone without steering their answer.
The AI recommends smaller portions
They hand the data to an AI to analyze. It proposes cutting portion sizes. The waste number drops. Done?
The number they weren’t measuring
Comparing similar menu days and talking with cafeteria staff raises the question the waste gauge can’t answer: would smaller portions cut waste because kids choose better, or just because some kids now leave hungry?
Let kids choose, with seconds available
New plan: each child picks a starting portion, and more food is available. Adults approve food safety and operations. The students run a limited trial.
Show the work, then get a new problem
The final work holds the calculations, interviews, decision, results, and what they’d change. Then a teacher hands them an unfamiliar graph and asks why comparing two different menus could mislead them.
A proposal can improve the measured number while making someone’s day worse.
Sound familiar? It’s the lesson of the testing era, layer eight of the trench coat, learned at the scale of a lunch tray. The difference is that these students learn it by owning a decision. The next time an AI recommendation lands in front of them, at a job or a city council meeting, it gets the same question: what isn’t this number measuring?
The project also pulls mathematics, science, writing, economics, and ethics into one piece of work. And it gives students a stable community while their academics move at their own pace: a child could need extra reading support while doing advanced science, keeping the same advisor and peer group while instructional groups shift as needed. Basic belonging is never earned by being good at things. Every child has a voice; readiness only decides which tasks they run on their own.
Two transcripts, three kinds of help
If the old receipts broke, print better ones. Start by naming which kind of help is happening, because each one serves a different purpose:
The student builds or shows an ability with no generative help. Assistive technology still counts as the student’s own: independence means real control and understanding, not denying anyone the support they need.
AI explains, asks questions, gives hints and feedback. The student does the thinking. (The guardrailed tutor lived here.)
The student uses powerful tools to get a real result, and shows how they directed and checked the work.
A student might learn to read a graph alone, get tutored on a misconception, then use AI to analyze a big dataset. Three modes, three purposes, all legitimate. Then report two kinds of capability separately. Tap through to see why.
Explains the ideas, handles a new problem, remembers it weeks later
Hesitant to direct tools; checks slowly
Shaky on the core ideas without help
Gets impressive results directing AI and teammates
Two very different students, one identical grade. The blend hides what each one needs next.
For the lunch project, “on your own” means a student can interpret a new graph, explain why comparing two different menus could mislead them, or rethink the plan after hearing an objection. Add delayed checks, because remembering and applying something weeks later matters more than performing right after the lesson. Keep common assessments so schools can still spot gaps and compare outcomes. A student who needs longer gets more time and teaching; a student who’s ready goes deeper. Neither becomes a permanent category of person.
What the day looks like
- MorningAdvisor group
A small group with one adult who knows you over years.
- Sustained reading, then math
Explicit teaching, worked examples, practice.
- Targeted practice
Tutoring or practice matched to what you need now.
- AfternoonLab, workshop, rehearsal, or project
Where the morning’s knowledge gets used.
- ProtectedMovement, meals, friends, play
Time that doesn’t have to produce anything useful.
Some days lean toward formal instruction; others leave room for long investigations or placements. The timetable serves the learning instead of the other way around.
And the adults matter more, not less. Every student keeps a continuing relationship with a human advisor. Subject specialists teach, diagnose misconceptions, and guide advanced work. AI can generate exercises, offer extra explanations, translate materials, and trim paperwork, but adults stay responsible for children’s welfare and for the judgments that shape their opportunities. Cheaper software is not a license for ever-bigger classes. Student records get limits too: no child should acquire a permanent algorithmic identity as “low potential” or “difficult,” and any automated recommendation about a child’s progress gets human review and a way to challenge it. The doors also stay open after graduation, because a world of fast-improving AI makes a one-time education less sufficient, not more.
Every layer measured something easier than learning
Walk down the trench coat one last time and look at what each fix could actually count. Compulsory attendance counted bodies in seats. The Carnegie Unit counted hours. The testing era counted scores. The homework folder counted products. Each was a reasonable stand-in for learning when it was invented, and each slowly became the thing schools were asked to produce.
AI is the first technology that can fake the stand-in at scale. That’s why it feels like a threat to school. It’s also the opening: it forces the question every layer let us skip.
America became remarkably good at getting children into school. The harder promise, making sure each one gets the teaching and support to learn well, is still unfinished. More access, more time, better teaching, fairer funding, and better measurement fix different failures, so the first step in judging any reform is naming which failure it’s meant to fix.
The redesign above is my answer to one of them. Underneath all of it is a single question worth asking of every kid, every year:
“Can this kid understand, decide, and do, with help and without it?”
That’s the receipt worth printing. Everything else in the room, the rows, the bell, the credit hours, can be rearranged around it. You can even keep the clock.
The room has changed five times. This time, change the grammar.
This essay began as a conversation with an AI about the history and future of American schooling, and was then checked against sources. Key sources: Tyack and Tobin, “The ‘Grammar’ of Schooling” (1994); Horace Mann’s Twelfth Annual Report (1848); Claudia Goldin’s research on the high-school movement; the Carnegie Foundation on the Carnegie Unit; the Interior Department’s Federal Indian Boarding School Initiative; GAO work on school funding and segregation; a National Research Council review of test-based incentives; Gallup on satisfaction with K–12 education, school performance by responsibility, students’ preparedness, and the importance of college (all 2026); the 2024 twelfth-grade NAEP results and 2025 long-term trend NAEP; the Education Department on chronic absenteeism; Battelle’s 2026 workforce survey; a Census Bureau working paper on business AI use; a 2026 national teacher survey on AI guidance; BLS earnings by education (2025); the National Academies on how people learn; UNESCO’s AI competency framework for students; Bastani et al. on AI tutors in high-school math; and a randomized study of an AI tutor in college physics. The toy model, the lunch project, and the two students are illustrations, not data.
If “performance with the tool is not ability without it” stuck with you
We Automated the Apprenticeship
The same receipt problem at work: when AI does the entry-level tasks, where does anyone learn the job?
Read the essay →Essay · Content PilotsThe Robot Dividend Could Hit $10 Million a Year
Job six, “a good life,” assumes work may take up less of it. Here’s how the money side of that future might work.
Read the essay →Essay · Content PilotsYour Data Is Perfectly Safe. That’s Exactly the Problem.
Why no child should get a permanent algorithmic label, and what a fair way back from an automated judgment looks like.
Read the essay →Primary source · AERJ, 1994The “Grammar” of Schooling
Tyack and Tobin’s classic on why the graded school, the Carnegie Unit, and their cousins outlast nearly every reform.
Read the paper →Teaching something? Make the front door as good as the lesson.
Tutors, coaches, music teachers, makerspaces, and training businesses are already running the kind of learning this essay describes. If that’s you, families should be able to find you, book a session, and pay without a week of phone tag. That’s what I build at Content Pilots: websites with scheduling and payments built in. And if you’re an educator who thinks I got the history wrong somewhere, email me. I’ll hand you the red pen.