1863 Leadership · Issue Paper No. 12
Borrowed Books
Lincoln learned law from books a neighbor lent him. Four repairs, and a library open to anyone on earth.
Abstract
American schools spend roughly $18,600 per student per year, four dollars in five of it on people — rightly so, since a school is adults in a room with children. What is spent on the material children actually learn from cannot be stated precisely, because no federal line item exists for it. What can be stated is that the money is small, that better materials cost essentially the same as worse ones, and that the market is sliced so finely across thirteen thousand districts and fifty state adoption cycles that no serious company can justify building for it. That is why educational software is bad. This paper proposes four repairs that leave schools, teachers and budgets alone: a free public learning platform open to anyone in the world, built to entertainment standards and running from true beginner through advanced with credentials employers recognize; a commitment to pay for demonstrated learning rather than for promises; universal connectivity funded by outcome rather than by technology; and the withdrawal of the federal government from student lending, which the evidence shows raises tuition about sixty cents on the dollar. We also set out, at length, the reforms we considered and rejected — vouchers and private school management — because the research on both runs against them.
Key findings
- 28 percent of American adults — 58.9 million people — read at the lowest measured level, up from 19 percent in 2017. For numeracy it is 34 percent, 72 million. Average literacy fell 12 points in six years.1
- The problem is mathematics, not education generally. U.S. PISA math is below the OECD average with 25 systems ahead of us and the lowest American score on record. Reading is above the OECD average with only five ahead.2
- Of $18,614 spent per pupil, salaries and benefits are 77.5 percent of operating costs. There is no federal line item for instructional materials at all — the nearest category lumps textbooks together with heating oil.3
- Federal student loan caps pass through to tuition at about 60 cents on the dollar for subsidized loans — and the increase lands on sticker price, so students who never borrow pay it too.4
- We reject vouchers on the evidence. In Louisiana the median voucher student fell from the 50th to the 34th percentile in math; third through fifth graders to the 26th. Effects across four studies run −0.15 to −0.50 standard deviations, among the largest declines in the research record.5
Section 1The question before us
A game studio is deciding whether to spend two hundred million dollars building something. The first question in the room is who pays it back. If the answer is thirteen thousand school districts, each with its own budget cycle, purchasing committee, pilot program and eighteen-month sales process, the studio builds another shooter instead — and the educational contract goes to a firm whose core competence is winning contracts.
That is why the software is bad. Not because nobody could build better. Because nobody serious can get paid for it.
The studios are not indifferent to children. They are looking at a spreadsheet.
This paper is narrow on purpose. It does not reorganize schools, redirect their budgets, or propose anything that costs a teacher a job. Seventy-eight percent of school spending is people, and a proposal about that money is a proposal about labor, which is a different argument than the one we are making. We are interested in the two percent nobody is competing for, and in what could be built alongside the schools rather than instead of them.
1.2 Our own share of the failure
People who argue for markets in education have spent thirty years arguing for one specific market — vouchers — and have been slow to read the results.
Section 8 sets them out in full. They are bad. Four rigorous studies using different methods found that students who used vouchers scored worse than similar students who stayed, with effects among the largest negative findings in the field. The argument has gone on being made anyway, and the honest reason is that the theory was attractive.
There is a second thing to own. The claim that America is falling behind the world in education is repeated constantly by people who agree with much else in this series, and it is not true as stated. In reading, American fifteen-year-olds score well above the OECD average and only five education systems on earth beat us. The collapse is in mathematics and it is real. Saying the larger thing because it sounds worse is how an argument loses the people who check.
Section 2What the money actually buys
Public schools spend about $18,614 per pupil per year. Of that, $16,280 is day-to-day operations, $1,833 is construction, and $501 is interest on school debt.3
Within operations, salaries and wages of $416.7 billion plus employee benefits of $178.3 billion account for 77.5 percent of current spending nationally. Instruction is 58.8 percent of the total and support services — which means counselors, nurses, librarians, custodians, bus drivers, cafeteria staff and principals, not a bureaucracy — is 36.2 percent.3
So roughly $14,400 of every $18,600 goes to people and $3,700 to buildings and operations. What remains for the material children actually learn from is small — and here we have to be careful, because the federal statistics will not tell you how small.
There is no line item for instructional materials. The closest federal category is "supplies," which is about 8 percent of district budgets and which the statistical agency describes as covering everything from textbooks to heating oil. Purchased curriculum is folded into instruction; curriculum development support sits in a separate category funded at a few hundred dollars per student; and technology spending is scattered across several categories at once.3
That a country spending $950 billion a year on schools cannot say what it spends on what children read is itself worth noticing.
What can be said precisely is the thing that matters more. Polikoff and Koedel put it exactly: textbooks are relatively inexpensive and similarly priced, so the marginal cost of choosing a more effective one over a less effective one is essentially zero. By one estimate, improving curriculum quality is roughly forty times more cost-effective than reducing class size.3
The content is the cheapest thing in the building and among the most consequential. Which makes it strange that it is the one thing nobody is competing to improve.
2.1 Why a ten-billion-dollar market behaves like a small one
Call the national market for instructional content ten to twenty billion dollars a year. That sounds enormous until you cut it the way a company selling into it must.
Thirteen thousand districts buying separately. Thirteen grade levels. A dozen subjects. Fifty states with different standards and staggered adoption cycles, so a district that bought last year is not buying again until 2029.
Slice it into pieces anyone can actually sell into and eighth-grade American history in one state's cycle might be worth a few tens of millions over several years, to be divided among incumbents who already have sales forces in those buildings.
A single major video game costs two to three hundred million dollars to build and market. The arithmetic explains everything about the quality of what schools are offered.
Section 3Reform I: A public library for the next century
The federal government should fund a free, permanent, online learning platform — built to entertainment production standards, running from true beginner through advanced, carrying credentials that employers and universities recognize, and open to anyone on earth.
Not a curriculum. Not a mandate. A library. Schools may use it as a course, as supplementary support, as a tutoring spine, or not at all, exactly as they use the library now.
3.1 What it covers
Literacy. Language. The sciences. Technology, engineering and mathematics. Medicine. Markets and business. Civil and criminal law. Structural civics — the branches of government, what a sheriff does that a district attorney does not, what the Declaration and the Constitution say. And history, handled differently, as Section 3.3 explains.
Every subject at three levels — beginner, intermediate, advanced — with a credential for completing each.
And it must be navigable. The failure mode of a library is a person who does not know where to begin and leaves. Three levels, a fixed list of subjects, an obvious front door.
3.2 The beginner tier is the whole argument
This is the part that does not exist anywhere, and the numbers explain why it matters most.
In 2023, 28 percent of American adults — 58.9 million people — scored at Level 1 or below in literacy, up from 19 percent and 48 million in 2017. In numeracy it is 34 percent and 72 million. Average literacy scores fell twelve points in six years, and the gap widened: the top held steady while the bottom grew.1
Level 1 means a person can read a simple sentence and struggles with multi-step instructions.
Khan Academy assumes you can read the problem. Fifty-nine million American adults cannot, and there is no front door for any of them.
A twenty-six-year-old reading at a fourth-grade level cannot start at Khan Academy, cannot start at Coursera, and is not going back to high school. They are not served by anything, at any price, and they are one in four adults.
This also answers the strongest objection to the whole idea, which Section 8.5 states at length: free learning tools are used most by the people who need them least. That is true of tools which assume prerequisites their users lack. A genuine beginner tier, ending in a credential with labor-market value, is a different product for a different person.
3.3 Civics yes, history as a collection, and no politics in either
We draw a line here deliberately, because the alternative is a permanent fight.
Structural civics can be taught without prescribing a single value. What the three branches are. The difference between a councilman, a mayor and a sheriff. Civil court versus criminal court. What powers a president has and what powers the office does not. What the Declaration says and what the Constitution says. None of that requires anyone to be told what to think, and the model already exists in nonpartisan civic education that has run for years without controversy.
Where constitutional interpretation is genuinely disputed — and it often is — the platform presents the dispute rather than resolving it.
History is a collection, not a curriculum. Books, films and programs representing multiple perspectives, assembled the way a library assembles a shelf. A person interested in the Civil War encounters several accounts of it and forms their own view.
We are not pretending this is neutral. Curation is editorial; somebody decides what is on the shelf. Libraries handle that with published selection standards and an independent professional board rather than by claiming to have no perspective, and this must do the same — with the author of the standard barred from choosing the titles.
One rationale we deliberately do not use: that this produces better voters. It might. But a government funding educational content on the ground that it will improve the electorate has announced that it intends to shape political opinion, and every administration after this one would rewrite it on that basis. Two rationales carry the proposal without inviting the fight — opportunity, and a workforce.
3.4 How it gets built, and for how much
Contracts are awarded by subject and age band, not as one master agreement. That is the concentration principle from Issue Paper No. 7: several firms compete, no one becomes the permanent incumbent, and the failure of one module does not take the platform down.
Each must be large enough to attract people who can actually make something children choose to use. If a single major game costs two to three hundred million, then twelve subject areas at that standard implies roughly $2 to $3 billion over five years.
Which sounds enormous until it is placed beside the schools. Total public K-12 spending runs near $950 billion a year. This is approximately five one-hundredths of one percent of it — about six dollars per student per year, for the entire country, permanently.
Two prohibitions. No advertising, ever, in any form: a federal platform serving children is the last place profiling belongs, and Issue Paper No. 6 explains why at length. And no exclusivity — existing providers like Khan Academy may and probably should win contracts, but no winner acquires a claim on renewal.
Finally, it is open to the world. The marginal cost of an additional user is approximately zero, scale makes the product better, and Abraham Lincoln taught himself law from borrowed books. This is the cheapest instrument of American influence ever devised, and we would rather say that plainly than pretend the motive is purely domestic.
Section 4Reform II: Pay for learning, not for promises
Alongside the platform, the government should commit in advance to purchase demonstrated learning gains at a stated price — naming no vendor, running no competition for a contract, and paying only on results verified by an independent party.
The mechanism is not novel. In 2009, five governments and the Gates Foundation committed $1.5 billion against pneumococcal disease, which was then killing more than seven hundred thousand children a year in poor countries. The terms were published in advance: meet this target product profile, commit to supply for ten years at no more than $3.50 a dose, and receive a share of the fund in proportion to what you supply. An independent assessment committee — not the donors — decided whether a product qualified.6
We should be precise about what it did and did not achieve, because the popular version overstates it. A 2021 evaluation found no evidence that the commitment accelerated the creation of vaccines; the underlying science largely existed. What it did was guarantee a market that manufacturers would otherwise have ignored, pull production capacity into serving it, and drive the price down — a later entrant supplied at $2.00 a dose, 43 percent below the starting price.6
That is the honest claim for education too. This would not invent a way of teaching nobody has thought of. It would make an existing capability — people who know how to build things children cannot stop using — worth pointing at a problem it currently ignores.
Applied here it means stating an outcome narrowly — say, algebra readiness for students in the bottom third — setting a bar, naming a price and a volume, publishing it, and then getting out of the way. Anyone competes. A third party randomly assigns students and measures. Everyone who clears the bar gets paid.
Procurement pays for a promise and picks the winner first. This pays for a result and picks nobody.
And it moves the risk. Under procurement the taxpayer carries the possibility that the product does not work. Here the developer carries it. If nothing clears the bar, the public pays nothing.
Section 5Reform III: Connectivity, funded by outcome
A library nobody can reach is not a library. Universal connectivity should be funded by defining the outcome — every household reaching a stated speed — and letting fiber, fixed wireless and satellite compete for each location on cost.
The precedent is the Rural Electrification Administration of 1936, which made power universal on the reasoning that it was infrastructure rather than luxury. The same argument applies now, and it belongs to the same American tradition as the Homestead Act and the GI Bill discussed in Issue Paper No. 11.
One technical correction, because the intuitive version of this proposal is wrong. It is tempting to say we need no new infrastructure because 5G already exists. 5G runs on fiber. Every cell site requires a fiber backhaul connection, and new fiber deployment is what makes wireless expansion possible. The wireless portion is the last mile only.7
Nor is coverage the same as capacity. Carriers prioritize mobile traffic over fixed, and deprioritize home users during congestion. A household streams far more than a phone.
That said, the deeper instinct is right and is supported by serious criticism of current policy: fixed wireless is faster and cheaper to deploy than trenching every street, and federal programs have been faulted for preferring fiber where wireless or satellite would reach more homes per dollar.7 Hence technology neutrality: fund the result, not the method.
Two honest qualifications. Satellite is the answer for genuinely remote homes and is not yet a general substitute — measured speeds have fallen below the federal standard in many states.7 And deployment and affordability are different problems: one is capital, the other a monthly bill. Universal deployment with targeted help on the bill is cheaper than universal free service and reaches the same households.
Section 6Reform IV: The federal government leaves student lending
Federal student lending should end on a date certain, three years after enactment. Pell grants remain. Research funding remains. The existing portfolio is repaid.
6.1 The same machine as health care
This is the argument of Issue Paper No. 10 in a different market, and it is worth naming the shape rather than treating student debt as a moral question about young people.
A payer with effectively unlimited capacity. A buyer who does not feel the price at the moment of purchase. A seller with no reason to restrain it. That arrangement produces the same result in medicine and in higher education, and for the same reason.
The evidence is unusually good. Research published in the Review of Financial Studies out of the Federal Reserve Bank of New York found a pass-through to tuition of about 60 cents on the dollar for increases in subsidized loan maximums and about 20 cents for unsubsidized, averaging near 40 cents across aid types. The effect was largest for expensive degrees, for-profits and two-year programs, and was corroborated by earnings-call transcripts and stock market reactions at publicly traded institutions. Because it lands on sticker price, it raises costs for students who never borrow.4
Notably, the effect on Pell grants was not statistically distinguishable — which is why grants stay and lending goes.
6.2 What happens to the loans that exist
Roughly $1.6 trillion is outstanding across some 43 million borrowers. It should be repaid.
Repayment terms may be extended and monthly payments scaled to income. But the distinction that matters is precise and must not be blurred: extended terms with eventual repayment in full is not forgiveness. The cost to the taxpayer of a longer term is the time value of money. The cost of discharge is the principal. We propose the first and not the second.
This also identifies where the budgetary savings actually are. Lending itself has often scored near break-even under federal credit accounting; what costs real money is discharge at the end of an income-driven schedule. Those are separable decisions and the paper separates them.
Section 7Two reforms we offer for debate rather than propose
Education is constitutionally the business of states, districts and parents. The following two ideas have evidence behind them and we think they deserve serious consideration. We are not proposing that anyone in Washington require them.
Mastery-based progression. Children are sorted by birth year, which is an administrative convenience rather than a theory of learning, and advanced by seat time measured in credit hours — a unit the foundation that created it is now trying to move past. The alternative is advancing on demonstrated mastery: the fifth grader ready for eighth-grade mathematics moves, and the struggling reader receives more time rather than a passing grade. It costs nothing. It requires permission.
High-dosage tutoring. It is the best-evidenced intervention in education. A meta-analysis of 89 randomized trials found a pooled effect of 0.288 standard deviations, roughly four additional months of learning, with the largest effects when tutors are teachers or paraprofessionals, sessions run at least three times weekly, and the work happens during the school day.8
With one caution that any policymaker must hear. An expanded meta-analysis of 265 trials found effects falling from 0.44 SD in programs under 100 students to 0.16 SD at a thousand or more. The scaling penalty survives every control.8 Tutoring works; tutoring at national scale works considerably less well, and nobody has solved that.
Section 8 · The strongest case against
8.1 Vouchers, which we considered and rejected
We began this work intending to propose a voucher system and the evidence stopped us.
Four rigorous studies using different designs — Louisiana, Indiana, Ohio and the District of Columbia — found that students who used vouchers performed worse than similar students who stayed. In Louisiana the median student fell from the 50th to the 34th percentile in mathematics after one year, and third through fifth graders to the 26th. Indiana: 50th to 44th. Ohio's evaluation reported overwhelming evidence of substantial negative effects. Magnitudes run −0.15 to −0.50 standard deviations, and Brookings noted that negative effects of any size are rare in education research.5 Louisiana's math effects persisted four years later.
The explanation matters more than the finding. The theory assumed private schools were better. Public schools have been under accountability pressure since the 1990s and private schools were exempt from it, and the gap closed. Louisiana's results were attributed to low-quality private schools the state approved — schools that participated because they had empty seats.5
There is a real finding on the other side and we report it. Three research teams studying Florida found that competition from its voucher program produced modest test-score gains in the traditional public schools.5 So the mechanism may work — on the children who never leave. A reader who weighs that against the direct effects differently than we do is making a defensible judgment.
We also note a structural risk independent of achievement. A school that loses half its students loses roughly half its revenue and nowhere near half its costs; the building, the heating and the principal do not shrink. The children who remain are the ones who bear that.
8.2 Private management of schools, also rejected
In 2002 Pennsylvania took over Philadelphia's 200,000-student district and handed 45 of its lowest-performing schools to seven private managers, including the nation's largest for-profit operator. The managers received additional per-pupil funding. It remains the largest experiment of its kind in American history.
RAND's evaluation found gains generally on par with comparison schools elsewhere in the state — no superior results despite the extra money — and found that one operator's converted schools fell behind comparison schools in the first year.9
The detail that settled it for us: the strongest results came from the district's own restructuring office, which received the same additional resources and ran the schools itself.9 The money helped. The privatization did not.
8.3 Screens are worse than paper for learning to read
Two meta-analyses settle this. Delgado and colleagues pooled 54 studies covering more than 170,000 participants and found screen reading worse by g = −0.21. Clinton's review of 33 randomized studies found g = −0.25.10
The important part is where the effect lives. Both found it almost entirely in expository text — the reading-to-learn material that fills a curriculum — and essentially absent in narrative. Delgado: −0.27 expository against +0.01 for narrative. Clinton: −0.32 against −0.04. Reading a novel on a screen appears to cost nothing. Reading to learn costs real comprehension.10
And it is growing rather than fading with familiarity: Delgado found the disadvantage increasing by about 0.01 each year from 2001 to 2017.10
One finding cuts the other way and belongs in any contract written under Section 3. A later network meta-analysis suggests paper's advantage may appear chiefly where the digital version requires scrolling.10 If that holds, the defect is a design choice rather than the medium, and it is fixable.
Pandemic remote instruction produced severe losses concentrated among the lowest-performing students. And PISA found students with under an hour of daily leisure device use at school scored about 50 points higher in mathematics than heavy users, with 65 percent reporting device distraction during math lessons.2
This is why nothing in this paper proposes replacing books, replacing teachers, or putting a device in front of a child during the school day. The platform is a library. A library is a place you choose to go.
8.4 Technology can produce the appearance of learning
The most important experiment in this area gave roughly a thousand high school students unrestricted access to a large language model during mathematics practice. Performance while assisted rose 48 percent. Unassisted exam performance fell 17 percent. They learned less while appearing to learn more.11
Set against that, the successes are instructive. A Harvard trial with 194 physics students found a purpose-built AI tutor outperformed in-class active learning — not lectures, but the research-backed pedagogy that beats lectures — with students learning more than twice as much in less time. A World Bank trial in Nigeria produced 0.23 standard deviations in English over six weeks, among the most cost-effective interventions measured.11
The difference is not the technology. The successes were built around specific objectives with scaffolding against shortcuts and teacher facilitation. The failure was general-purpose access. Winning must require the reasoning, not merely permit it — and any contract under Section 3 must be written to that standard, with payment contingent on unassisted performance.
8.5 Free tools are used by the people who need them least
This is the strongest objection to the platform and we have only a partial answer.
Voluntary learning resources are taken up disproportionately by motivated students with engaged parents and quiet places to work. Access that is free to everyone is used unevenly, which means a purely voluntary platform can widen the gap it was built to close.
Three things push against it: the beginner tier serves a population currently served by nothing, credentials with labor-market value give an adult a concrete reason to finish, and schools and tutoring programs can build on it rather than leaving uptake to chance. None of that guarantees the result, and a reader who thinks availability is the easy part has understood the problem correctly.
8.6 Much of this already exists
Khan Academy is free, comprehensive and used by well over a hundred million people. So are CK-12, PBS LearningMedia and Duolingo. The honest case for a public version is production quality, the beginner tier, and permanence — philanthropy funds Khan Academy and philanthropy can stop. A reader who concludes the field is adequately served without federal money is making a reasonable argument, and our answer is that the fifty-nine million adults in Section 3.2 are evidence that it is not.
8.7 Ending student lending hits students before it lowers prices
The research in Section 6.1 finds that sixty cents of each new subsidized dollar went into tuition. It does not follow that withdrawing the dollar removes sixty cents of price next semester. Universities carry tenure obligations, bond payments and long cost structures; the adjustment could take a decade, and during it enrollment falls before prices do.
Private lenders would return and would price by risk — by institution, by field of study, by credit and cosigner. A nursing student at a state university gets a rate. A first-generation student with no cosigner studying history may not get an offer. That is what risk pricing does, and it falls hardest on precisely the students the rest of this paper is built to serve.
8.8 What we concede, and what we do not
We concede that screens are worse than paper for comprehension and that educational technology has a long record of disappointing. We concede that free access does not produce equal use. We concede that much of the platform exists in some form already. We concede that the advance market commitment may be too small to move a major studio, and that curation of a history collection is editorial however carefully it is governed. We concede that ending federal lending will hurt some students before it helps any.
We do not concede that the present arrangement is a market. Three hundred dollars per student per year, sliced across thirteen thousand districts and fifty adoption calendars, purchased by committees that will never meet the child, is not a market that failed. It is a market that was never permitted to form.
Section 9What we are not claiming
We are not claiming this fixes schools. It does not touch them. Seventy-eight percent of school spending is people, and nothing here redirects a dollar of it.
We are not claiming technology teaches better than teachers. The evidence says the opposite in most conditions, and Section 8.4 sets out the one experiment that should make any enthusiast careful.
We are not claiming America is behind the world in education. In reading and science it is ahead of the developed-world average. In mathematics it is behind and falling, and among adults the literacy numbers are a genuine emergency.
And we are not claiming to know this would work. It has not been tried at this scale or to this standard. By the test this organization applies to everyone else, that is a real objection, and the honest reply is only that it costs five one-hundredths of one percent of what the country already spends and takes nothing away from anyone.
Section 10The argument you can carry
The paper compressed to what a person can remember and repeat.
And the one that indicts the argument's own side. People who argue for markets in education spent thirty years arguing for one specific market and were slow to read the results. The evidence on vouchers has been in for a decade. The claim kept being made anyway, because the theory belonged to the people making it.
Section 11Conclusion
Abraham Lincoln had about a year of formal schooling. He learned law by borrowing books from a neighbor and reading them by firelight, which is a story Americans tell fondly and rarely examine. What it describes is a country where the only thing standing between a poor boy and an education was access to the material.
We have solved that problem for everyone with a fast connection, a quiet room, and the reading level to begin. Tens of millions of American adults do not have the third one, and nothing currently sold or given away starts where they would have to start.
Nothing in this paper reorganizes a school, redirects a budget, or threatens a teacher. It proposes building something alongside the system rather than in place of it — free, permanent, beginning at the true beginning, ending in a credential somebody will hire you for, and available to anyone in the world who wants it.
Against $950 billion a year, it costs about six dollars per student. Against the alternative, which is one adult in four unable to read a set of instructions in the wealthiest country on earth, it is not a close question.
In 1863 the country gave land to anyone who would work it. It cost the Treasury almost nothing and it built the middle of America. The thing we have in surplus now is not land. It is knowledge, and the cost of giving it away is the cost of a server.
Notes
- Program for the International Assessment of Adult Competencies (PIAAC), Cycle 2, conducted 2022–2023; U.S. results released by the National Center for Education Statistics in December 2024, nces.ed.gov. Summary figures — 28 percent and 58.9 million adults at Level 1 or below in literacy, up from 19 percent and 48 million in 2017; 34 percent and 72 million in numeracy; a 12-point decline in average literacy — as compiled by ProLiteracy, proliteracy.org. ProLiteracy advocates for adult literacy funding; the underlying data is OECD and NCES. Note that NCES cautions against direct comparison with earlier cycles owing to changes in administration and scoring; the 2023 assessment was delivered entirely on tablet.
- OECD Programme for International Student Assessment, 2022 results, United States country note, oecd.org, and NCES, nces.ed.gov, the source for the counts of systems outscoring the United States — five in reading, twenty-five in mathematics, nine in science — and for U.S. scores of 504 in reading against an OECD average of 476. On the device-use association, see the PISA 2022 findings as discussed by the National Education Association, nea.org; the NEA opposes several reforms in this series and the association is correlational rather than causal.
- National Center for Education Statistics, Fast Facts: Expenditures, nces.ed.gov, for the 2020–21 per-pupil figures of $18,614 total, $16,280 current, $1,833 capital outlay and $501 interest; and NCES, "Revenues and Expenditures for Public Elementary and Secondary Education," nces.ed.gov, for salaries and wages of $416.7 billion plus benefits of $178.3 billion equalling 77.5 percent of current expenditures. Instruction and support-service shares are from the U.S. Census Bureau's FY 2024 school system finances release, census.gov. On the absence of an instructional-materials line item: NCES reports a combined "supplies" object amounting to roughly 8 percent of district budgets and covering, in the agency's own description, everything from textbooks to heating oil; purchased curriculum is recorded under instruction while curriculum development support falls under instructional staff services, and technology spending is distributed across several categories. See EdWeek Market Brief, "K-12 Spending: Where the Money Goes," marketbrief.edweek.org, reporting NCES's own account of its categories. On the marginal cost of better materials, Morgan Polikoff and Cory Koedel, quoted in EdReports, "Why Materials Matter," edreports.org, which is also the source for the estimate that improving curriculum quality is roughly forty times more cost-effective than class-size reduction. EdReports advocates for higher-quality instructional materials and readers should weigh its framing; the Polikoff and Koedel observation should be cited to their own published work before publication.
- David O. Lucca, Taylor Nadauld and Karen Shen, "Credit Supply and the Rise in College Tuition: Evidence from the Expansion in Federal Student Aid Programs," Review of Financial Studies 32(2): 423–466 (2019), originally Federal Reserve Bank of New York Staff Report 733, newyorkfed.org. Source for the 60-cent subsidized and 20-cent unsubsidized pass-through, the roughly 40-cent average across aid types, the concentration of the effect in expensive degrees and for-profit and two-year institutions, the corroborating evidence from earnings calls and equity prices, and the finding that Pell effects were not statistically distinguishable. The general proposition that federal aid increases tuition is known in the literature as the Bennett Hypothesis, after a 1987 essay by Education Secretary William Bennett; this paper relies on the empirical work above rather than on the hypothesis as originally stated.
- On Louisiana: Abdulkadiroğlu, Pathak and Walters, and subsequent evaluations of the Louisiana Scholarship Program. On Indiana, Ohio and the District of Columbia, and for the synthesis of effect magnitudes and the observation that negative effects are rare in education research, see Brookings Institution coverage of the voucher literature, brookings.edu. On competitive effects in Florida, three independent research teams have found modest positive effects on traditional public schools. These literatures are contested and several of the organizations publishing on them advocate for or against choice; we have cited the direction and magnitude of findings rather than any organization's interpretation, and any published version should cite each underlying study directly.
- Michael Kremer, Jonathan D. Levin and Christopher M. Snyder, "Advance Market Commitments: Insights from Theory and Experience," NBER Working Paper 26775, summarized at nber.org. On mechanics — the $1.5 billion committed by Italy, the United Kingdom, Canada, Russia, Norway and the Gates Foundation; the ten-year supply commitment at no more than $3.50 per dose; the allocation of funds in proportion to supply; and the Independent Assessment Committee's role in certifying products against a Target Product Profile — see Gavi, gavi.org. On the limits of the result: a 2021 Dalberg evaluation did not find evidence that the commitment accelerated vaccine development, and Médecins Sans Frontières has published a critique of the mechanism, msfaccess.org. We report those objections because they bound what this paper claims: the commitment created a market and lowered a price; it did not invent a vaccine.
- On fiber backhaul requirements for 5G, capacity constraints and the deprioritization of fixed wireless traffic during congestion, and on criticism of federal technology preferences: Congressional Research Service, "The Broadband Equity, Access, and Deployment (BEAD) Program," R48666, congress.gov, which is also the source for measured satellite speeds falling below the 100/20 Mbps standard; and Information Technology and Innovation Foundation, "BEAD Needs All Technologies to Succeed," itif.org. ITIF advocates for technology neutrality, which is the position this paper adopts; readers should weigh the source accordingly.
- Andre Nickow, Philip Oreopoulos and Vincent Quan, "The Promise of Tutoring for PreK–12 Learning: A Systematic Review and Meta-Analysis of the Experimental Evidence," American Educational Research Journal (2024), journals.sagepub.com, for the pooled 0.288 SD effect across 89 randomized trials. On the scaling penalty — 0.44 SD below 100 students falling to 0.16 SD at 1,000 or more — see the expanded meta-analysis of 265 randomized trials, edworkingpapers.com.
- Brian P. Gill, Ron Zimmer, Jolley Bruce Christman and Suzanne Blanc, State Takeover, School Restructuring, Private Management, and Student Achievement in Philadelphia, RAND Corporation and Research for Action (2007), rand.org. Source for the 45 privately managed schools, the seven private managers, the additional per-pupil funding they received, the finding of gains on par with comparison schools, and the relative performance of the district's own Office of Restructured Schools. On Edison conversion schools falling behind comparison schools in the first year, see RAND, rand.org.
- Pablo Delgado, Cristina Vargas, Rakefet Ackerman and Ladislao Salmerón, "Don't throw away your printed books: A meta-analysis on the effects of reading media on reading comprehension," Educational Research Review 25 (2018): 23–38 — 54 studies, more than 170,000 participants, overall Hedges' g = −0.21, expository g = −0.27 against narrative g = +0.01, with the effect significant under time constraints and increasing by approximately 0.01 per year across 2001–2017. And Virginia Clinton (2019), a review of 33 independent randomized studies covering 2,799 participants, overall g = −0.25, expository g = −0.32 against narrative g = −0.04; see the summary at edweek.org. Both meta-analyses drew predominantly on undergraduate samples, which limits generalization to younger children. On the suggestion that paper's advantage may be concentrated where the digital version requires scrolling, see the subsequent network meta-analysis literature; that finding is more recent and less settled, and we report it as a possibility rather than a conclusion.
- On unrestricted generative AI access reducing unassisted performance: Bastani et al. (2025), randomized trial of approximately 1,000 high school mathematics students, reporting a 48 percent improvement in assisted performance and a 17 percent decline in unassisted exam performance relative to control. On purpose-built AI tutoring outperforming active learning: Kestin, Miller, Klales et al., "AI tutoring outperforms in-class active learning," Scientific Reports (June 2025), nature.com, a randomized crossover trial with 194 undergraduate physics students. On Nigeria: De Simone, Tiberti, Barron Rodriguez et al., "From Chalkboards to Chatbots," World Bank Policy Research Working Paper 11125 (May 2025), worldbank.org, reporting 0.23 SD in English over six weeks. The Harvard trial used a tutor built by the instructor around that course's exact objectives, which limits how far the result generalizes.
A note on the author
Issue papers are published under the name of 1863 Leadership rather than an individual byline.
A note on sources
This paper cites more evidence against its author's original instincts than any other in this series. We set out to propose vouchers and the research in note 5 stopped us; we set out to propose privately managed schools and the Philadelphia evaluation in note 9 stopped us; we set out to propose replacing books with screens and the literature in note 10 stopped us. Those findings are reported at full strength in Section 8 rather than buried. Note 3 flags our own estimate of instructional-materials spending as requiring proper sourcing, and notes 6 and 10 flag citations still to be obtained directly.
Recommended citation
1863 Leadership. "Borrowed Books." Issue Paper No. 12. September 2026. 1863leadership.org
Corrections: None to date. Errors of fact are corrected on this page within one business day of notice, with a dated note describing the change.