The Student Debt Machine — 23% of Degrees Lose Money

The central finding of this report is that American higher-education finance no longer prices the thing it sells. Outstanding student loan balances stood at 1.651 trillion dollars in the second quarter of 2026, inside a household debt stock of 18.771 trillion dollars [1]. Against that balance, roughly 23% of bachelor's degree programmes in the United States deliver a negative lifetime return — their graduates end up financially worse off than they would have been with a high-school diploma alone [3]. Both numbers are produced by the same machinery: a lender that does not underwrite, an institution that does not carry the loss, and a borrower who cannot discharge the debt. The result is not primarily an education problem. It is a credit market in which the only participant without a hedge is the eighteen-year-old signing the note.

The headline delinquency figure for 2026 looks deceptively benign. The New York Fed recorded 7.83% of student loan balances 90 or more days delinquent in the second quarter of 2026, down from 12.88% a year earlier [1]. The improvement is an artefact. The Fed attributes the decline to the continued effect of re-reporting previously defaulted debt into the credit files, not to borrowers resuming payment [1]. Reading that series as recovery is the single most common error in current coverage of the subject, and it matters because the underlying default numbers moved the other way.

Those numbers are stark. As of January 2026, 8.8 million borrowers were in default on loans held directly by the Education Department, covering 208.7 billion dollars of balances [2]. Of those, 3.62 million defaulted for the first time between 20 January and 31 December 2025, against effectively zero new defaults in the first quarter of that year [2]. Defaulted balances across the whole portfolio reached a record 233 billion dollars by mid-2026, with roughly one in five borrowers more than nine months behind [44]. A separate count put 8.931 million federal borrowers at least 30 days delinquent in the first quarter of 2026 [40].

This was structurally foreseeable. Federal payments were suspended for four years, then restarted, then covered by an on-ramp period during which missed payments carried no default consequence. When the on-ramp expired, several years of accumulated non-payment converged on the 270-day default threshold at once — a mechanical result the Congressional Research Service modelled in advance [31]. The wave was not a surprise about borrower behaviour. It was an accounting deadline arriving.

Set that against how any other credit market functions. A mortgage lender values the asset, checks the borrower's income and refuses the loan when the numbers fail. An auto lender prices the risk into the rate. In federal student lending, the loan amount is determined by the cost of attendance the institution declares, the rate is set by statute, and no underwriting of the borrower's expected earnings occurs at any point. The institution receives the money at enrolment and keeps it whether the student graduates, whether the graduate finds work, and whether that work pays enough to service the debt.

That design was defensible while the return on a degree was rising and its dispersion was narrow. Neither condition now holds. The remainder of this report traces what happened to the price, what happened to the premium, and what the programme-level data shows about the distance between them.

The economic case for student borrowing has always rested on a single empirical claim: that the wage gap between degree holders and high-school graduates is large and growing. For three decades that claim was accurate. It is no longer. Minneapolis Fed analysis puts the overall college wage premium at 55.2%, a level last seen in the late 1990s, after a peak around 2015 and a slow decline since [5]. A San Francisco Fed working paper reaches the same place by a different route, finding the gap essentially flat over twenty years and slightly below its 2000 value by 2023 [5].

Meanwhile the price kept moving. The same analysis notes that a four-year degree in 2022-23 cost 40% more in real terms than in 2000-01 [5]. Published prices for 2025-26 run to 11,950 dollars in tuition and fees for in-state students at public four-year institutions and 45,000 dollars at private nonprofit institutions [7]. An investment whose price rises 40% while its return flattens is not a marginal case. It is a repricing that the finance system around it never registered.

The honest complication is that published prices are not what most students pay. After adjusting for inflation, average net tuition and fees paid by first-time full-time in-state students at public four-year institutions peaked at 4,450 dollars in 2012-13 and had fallen to about 2,300 dollars by 2025-26 [7]. At private nonprofit institutions, real net tuition fell from 19,810 dollars in 2006-07 to about 16,910 dollars [7]. Anyone arguing that American undergraduate tuition is the proximate cause of the debt stock has to explain a series that has been falling in real terms for more than a decade.

That brings the discussion to where price pressure actually comes from. The Bennett hypothesis holds that expanding federal aid allows institutions to raise prices by roughly the amount of the expansion. The strongest quantitative test, a New York Fed staff report, found average pass-through of about 40 cents on the dollar, with wide variation by aid type [8]. For-profit institutions captured about 57 cents of each additional grant dollar and 51 cents of each additional loan dollar [8]. In that sector the mechanism is not disputed, only its magnitude elsewhere.

Where the money goes once it arrives is a separate and better-documented matter. Between 2010 and 2018, spending on student services rose 29% and administrative spending 19%, against 17% for instruction [46]. Over a longer window, instruction fell from 41% to 29% of total university spending [46]. Earlier work covering 1993 to 2007 found full-time administrators per 100 students up 39% while teaching, research and service staff rose 18%, with real administrative spending per student up 61% against 39% for instruction [45]. The institution grew around the teaching, not through it.

Hold the two series side by side. A premium that stopped growing a decade ago, a real price that fell at public undergraduate level but rose sharply in graduate study, and a lending facility indifferent to both. The aggregate case for borrowing has not collapsed. It has simply stopped being an aggregate case at all, which is what the programme-level data shows next.

The most consequential piece of research in this field is not about degrees as a category. It is Preston Cooper's programme-level return-on-investment work at the Foundation for Research on Equal Opportunity, which estimates lifetime financial return for roughly 53,000 individual degree and certificate programmes rather than for institutions or for higher education as a whole [3]. Adjusted for the probability that a given entrant does not finish, the median bachelor's programme returns about 160,000 dollars over a lifetime [3]. That is a real and substantial number. It is also an average across a distribution so wide that the average carries almost no information for an individual student.

The dispersion by field is the finding. Engineering programmes return a median of about 949,000 dollars, computer science about 652,000 dollars and nursing about 619,000 dollars [3]. Fine arts programmes return less than nothing, and education, English and psychology cluster near or below the line [3]. The gap between the top and bottom of that range is several times larger than the gap between holding a bachelor's degree and holding none. A financing system that lends identical amounts on identical terms across that range is not pricing risk badly. It is not pricing risk at all.

Graduate study is where the arithmetic turns hostile fastest. The typical bachelor's graduate holds about 28,000 dollars in federal loans; the typical master's graduate holds more than 55,000 dollars [11]. Around 40% of master's programmes do not raise earnings enough to justify their tuition [11]. Engineering, computer science and nursing master's degrees almost always pay; arts and humanities programmes rarely do, and even the MBA — one of the most popular graduate credentials in the country — frequently returns little or nothing [11].

The Urban Institute reached the same conclusion from the debt side rather than the earnings side. Master's and doctoral programmes in psychology, master's programmes in visual and performing arts, and for-profit graduate programmes in family and consumer sciences leave students with unaffordable debt or low earnings at the highest rates of any category examined [12]. Until July 2026, every one of those programmes could borrow without limit through Grad PLUS, at the full cost of attendance the institution chose to declare [13].

This is the point at which the public argument and the evidence part company. The recurring question in political discourse is whether college is worth it. The programme-level data says the question is malformed: it aggregates over a distribution in which the variance between programmes exceeds the mean difference the question is asking about. The correct question is whether a specific programme at a specific institution at a specific net price is worth it, and until very recently no part of the American financing system was organised to ask it.

Two objections deserve to be taken seriously, and both are addressed later in this report. The first is that a purely financial return misstates the value of education. The second is that Georgetown's lifetime-earnings modelling puts the bachelor's premium near 1.2 million dollars, an order of magnitude above the FREOPP median [10]. Neither objection touches the dispersion finding, which is what the lending system is blind to.

Credential inflation is usually described as a cultural drift. It is better understood as a rational response to an abundant supply of graduates. When the share of applicants holding a degree rises, a degree requirement becomes a cheap first-pass filter regardless of whether the work requires it. Burning Glass Institute analysis found that 43% of postings requiring a bachelor's degree described roles performable by workers with alternative credentials or relevant experience [9]. The requirement is doing screening work, not skill work.

Between 2017 and 2024 a large number of American employers publicly removed degree requirements, and roughly 20% of postings dropped them [9]. The follow-through was close to nil. Research by the Burning Glass Institute with Harvard Business School tracked what actually happened to hiring after the announcements and found that fewer than one in 700 hires changed as a result [9]. About 37% of firms that dropped requirements genuinely followed through, raising their share of workers without degrees by nearly 20%; the rest changed the posting and nothing else [9].

The labour market outcome of that ratchet is visible in two series. In the second quarter of 2026 the unemployment rate for recent American graduates stayed elevated at about 5.6% while the underemployment rate — the share working in jobs that do not require a degree — edged up to 42% [4]. Cleveland Fed work shows the young graduate unemployment rate has recently surpassed the overall rate, and that the erosion of the graduate unemployment advantage has been building since 1979 through falling job-finding rates rather than rising job loss [6].

The same mechanism operates on a far larger scale in China. A record 12.7 million university students graduated in summer 2026, about 4% more than the 12.22 million of the previous year — an additional 480,000 entrants into a labour market already slowing [25]. Youth unemployment reached 18.9%, the highest since the series was redefined in 2023 to exclude students [26]. The mismatch between what the graduates were trained for and what the vacancies require is pushing large numbers into blue-collar and gig work [26].

Korea shows the cost of the ratchet displaced onto households rather than onto debt. Korean families spent just under 29.2 trillion won on private education in 2024, about 20.2 billion US dollars, up 60.1% in a decade, with participation among school students running at roughly 78.3% [27]. That spending buys position in a queue for university places whose economic value depends on other families spending less — which they do not. The expenditure is real, the aggregate positional gain is close to zero, and none of it appears in any student debt statistic.

Put the American and Asian versions together and the structural picture is the same in both. Expanding the supply of credentials does not expand the supply of the jobs the credentials were meant to unlock. It raises the credential requirement for the jobs that exist. The individual return to getting the degree stays positive — which is why students keep enrolling and why they are not making a mistake — while the aggregate return to everyone getting it falls toward the cost of acquiring it.

Follow the money through a single transaction. A student enrols in a programme, borrows the declared cost of attendance, and pays the institution. The institution books the revenue immediately and has no further financial exposure to whether the student graduates, what the graduate earns, or whether the loan is repaid. The lender is the federal government, which does not mark its loan book to market and cannot become insolvent. The only party whose balance sheet is exposed to the outcome is the student, and that exposure is not dischargeable.

That last point is a statutory construction with a traceable history. Federal student loans became subject to an undue hardship standard without any waiting period in 1998, and the 2005 bankruptcy act extended non-dischargeability to qualified private education loans [47]. In practice, the undue hardship test is met rarely enough that most borrowers never attempt it. The obligation therefore survives unemployment, illness, divorce, business failure and the closure of the institution that issued the credential.

A large share of that exposure sits with parents rather than students. Parent PLUS loans carried no meaningful underwriting and no borrowing cap until the 2025 statute imposed one, and they are fully subject to Treasury offset against the borrowing parent's benefits in default [13][35]. From 2026 new Parent PLUS borrowers also lose access to income-driven repayment entirely [13]. The instrument that allowed a family to finance a child's degree without regard to the family's capacity to repay now has the tightest terms in the system.

The sector where these incentives produce the worst outcomes is well identified. New York Fed research found that students at for-profit institutions borrow more, have worse labour market outcomes and default at far higher rates than comparable students at similarly selective public institutions [30]. Historical cohort data shows more than half of for-profit borrowers entering repayment in 2003-04 defaulted within twelve years, roughly triple the rate at four-year public or private nonprofit institutions [30]. In 2026, about a third of for-profit borrowers were 90 or more days behind [40].

The downstream effects are measurable and fall on exactly the life decisions the credential was supposed to enable. Research in the Journal of Labor Economics finds that a 1,000 dollar increase in student debt lowers the homeownership rate by about 1.8 percentage points for public four-year college-goers in their mid-twenties, equivalent to an average delay of roughly four months [43]. Work using the repayment pause as a natural experiment finds student debt suppressing marriage formation and childbearing, with the effect appearing when payments stop and returning when they restart [42].

None of this describes a market failure in the usual sense. Every actor is behaving rationally within the rules. The institution maximises enrolment because enrolment is revenue. The employer raises the credential bar because the filter is free. The student borrows because the individual return is still positive. The government lends because access is the policy goal. The losses are real, they are large, and the design ensures they accumulate on the one balance sheet with no mechanism for discharging them.

Australia has run the most direct experiment. From 1 June 2025 the government applied a one-off 20% reduction to every eligible student loan balance, removing more than 16 billion Australian dollars of debt across more than 3 million borrowers, applied automatically by the tax office with no application required [16]. Indexation — 3.2% for 2025 — then applied to the reduced balance rather than the original [17]. The compulsory repayment threshold rose from 54,435 to 67,000 Australian dollars, and repayment became marginal, charged only on income above the threshold rather than on the whole of it [17].

The Australian design is worth understanding precisely because it changes the risk allocation rather than the price. HELP debt is income-contingent: nothing is owed below the threshold, the balance is indexed rather than charged interest, and the obligation dies with the borrower. That makes it closer to a graduate tax with a cap than to a loan. The 20% cut was a transfer, and a large one, but the structural protection was already in place before it — which is the part the American system lacks and the part no amount of cancellation would supply.

England took the opposite route: a high nominal price funded by an income-contingent loan whose subsidy is hidden in the accounting. Fees rose from 9,250 to 9,535 pounds in 2025-26, the first increase since 2017, and the cap is set to rise to 9,790 pounds in 2026-27 and 10,050 pounds in 2027-28 [18]. The Department for Education puts the Plan 5 RAB charge — the share of new lending never expected to be repaid — at 30% [19]. On that measure nearly a third of the loan book is a grant that has been labelled a loan.

That figure is disputed in a way that illustrates how little anyone knows. The Institute for Fiscal Studies puts the same 2025 entry cohort at minus 6%, implying a long-run taxpayer gain of 1.3 billion pounds, with the genuine grant element confined to a 10% write-off share worth 2.1 billion pounds [19]. The entire divergence rests on assumptions about graduate earnings in the middle of a 40-year repayment window [39]. Two competent institutions looking at the same cohort disagree about whether the state makes money or loses a third of its outlay, and neither can be checked for decades [20].

The Nordic and German systems remove the price question and reveal what sits underneath it. Germany charges no tuition in almost every state while spending about 17,960 US dollars per student at purchasing power parity across primary to tertiary education, above the OECD average of 15,023 [24]. Norway spends over 33,000 US dollars per student in public institutions [21]. And yet the OECD records that more than half of students in Norway, Sweden and Finland still borrow — because the binding constraint is the cost of living during study, not the fee [21]. Abolishing tuition does not abolish student debt. It changes what the debt is for.

Free tuition also does not solve completion, which is where the financial damage concentrates. Norwegian bachelor's completion for pandemic-era entrants rose four percentage points to 53%, with a 12-point gender gap — 81% of women against 70% of men [22]. Finland's gender gap runs at 17 points, 84% against 67%, well above the OECD average [23]. A student who borrows for living costs and does not finish carries the debt without the credential, and that is the single most reliable route into default in every system examined here.

Economists offer two explanations for why graduates earn more. Human capital theory says education makes people more productive. Signalling theory says education identifies people who were already more productive, and the credential transmits that information to employers. Bryan Caplan puts the signalling share at 45% on a cautious reading and 80% on his preferred reading, resting the argument on the sheepskin effect: the earnings jump at graduation is far larger than the return to the years of study preceding it [36].

The methodological rejoinder is stronger than most participants acknowledge. A formal analysis in Empirical Economics argues that human capital and signalling generate observationally equivalent wage patterns, so the relative share cannot be identified from earnings data alone [37]. Critics also contest Caplan's arithmetic directly, noting that his own estimate of ability bias at 25% to 45% of the premium is difficult to reconcile with an 80% signalling share [36]. The dispute is not close to resolution and may not be resolvable at all.

The same unresolvability afflicts the price debate. The Bennett hypothesis has strong support in one sector and weak support in another. A New York Fed staff report finds average pass-through of about 40 cents on the dollar, rising to 57 cents for grant aid at for-profit institutions [8]. A study of law school pricing — a sector with historically uncapped federal lending, where the hypothesis predicts its largest effect — found no strong evidence that schools raised prices in response [38]. Real net tuition at public four-year institutions has meanwhile fallen for more than a decade [7].

Public opinion is moving in a direction that fits none of the competing models neatly. Gallup recorded confidence in higher education rising to 42% in June 2025 from 36%, the first upward movement since tracking began in 2015, with those reporting little or no confidence falling from 32% to 23% [28]. At the same time 75% of current students and 71% of graduates say the degree is worth the cost [29]. The people inside the transaction rate it considerably higher than the public rates the institution — which is what one would expect if the individual return remains positive while the aggregate bargain deteriorates.

What would settle any of this is programme-level earnings data matched to programme-level debt, published as a matter of course. The United States has just begun building exactly that. The 2026 earnings-accountability framework would restrict Direct Loan eligibility to programmes whose graduates clear defined earnings benchmarks, replacing the narrower gainful employment architecture with a system-wide test [14][15]. Whether it survives contact with litigation and lobbying is a separate question from whether the measurement is the right one.

It is worth being precise about what the disagreement does not cover. Nobody disputes the size of the outstanding balance, the number of borrowers in default, the underemployment rate, or the dispersion of returns across programmes. The contested territory is causal and forward-looking: why prices rose, what the premium will do next, and how much of the return is skill rather than sorting. The descriptive facts are not in dispute, and they are sufficient to support the conclusions that follow.

The first conclusion is that the aggregate case for higher education survives the evidence. The median programme returns a substantial positive figure, graduates still earn a large premium over non-graduates, and the people who hold degrees overwhelmingly say they were worth the cost [3][10][29]. Any argument that college as a category is a bad deal has to explain those three facts, and none of the available critiques does.

The second is that the aggregate case has stopped being the relevant case. When 23% of bachelor's programmes, 43% of associate programmes and close to half of master's programmes return less than nothing, and when about 31% of students sit inside that band, the distribution matters more than the mean [3]. The financing system lends the same amount on the same terms across the whole of it, which converts a dispersion problem into a default problem with a predictable lag.

The third is that the credential requirement is now largely independent of the work. 43% of postings demanding a degree describe roles that do not need one, and the employer movement to drop requirements changed fewer than one in 700 hires [9]. This is a coordination failure with no market solution, because the firm that drops the filter first bears the cost of a worse screen while the benefit accrues to the labour market as a whole.

The fourth is that debt relief and price discipline are different instruments and are being confused for each other. Australia removed more than 16 billion Australian dollars of balances and changed nothing about what an institution may charge or what it bears when a graduate cannot repay [16]. England has moved the price up while writing off, on the official estimate, close to a third of what it lends [18][19]. Neither country has attached the loss to the party that sets the price.

The fifth conclusion concerns the countries that removed price from the equation, and it is the least intuitive finding in this report. Germany and the Nordic systems charge no tuition and still see most students borrow, because living costs bind where fees do not [21][24]. Completion rates in those systems sit near or below 53% for recent cohorts with gender gaps of 12 to 17 percentage points [22][23]. Free tuition changes who pays and what the debt is for. It does not by itself deliver the outcome that non-completion destroys.

What the evidence will not support is the claim that students are making a mistake. Given the terms they face, enrolling is usually rational, borrowing is usually rational, and choosing a high-return field is not available to everyone who would prefer it. The mistake is structural, it is located upstream of the student, and it consists of running a 1.65 trillion dollar credit market in which the only party who can lose is the only party with no information, no pricing power and no exit [1][47].