Education

A system under stress in nearly every developed country. Test scores falling since before the pandemic, the cost of college, the future-of-work mismatch, AI in the classroom, and what specific things help.

A note on framing. The page below tries to walk through what the picture actually looks like - what is working, what is not, what the post-pandemic period exposed, and what the AI era is starting to change - in a way that respects readers across the political spectrum. There is no clean ideological answer to most education questions, and the data on what specifically works is more useful than the slogans suggest.

The post-pandemic learning loss

The COVID-19 pandemic interrupted in-person schooling worldwide for periods ranging from a few months to over a year, depending on country and grade level. By 2022 the average US student had lost about half a year of math learning and a quarter of a year in reading, compared with a normal period. Similar patterns appear in British, Australian, Brazilian, and other data, varying with how long schools were closed and what took their place.

What worsened the picture: the loss was sharply unequal. Students from lower-income families, in lower-income school districts, and in countries with longer school closures lost more than their peers. Students whose parents could supplement schooling at home with paid tutors, technology access, or one parent staying home with the children fared substantially better. The pandemic widened the achievement gap that already existed, in some cases reversing decades of slow progress on closing it.

The recovery has been slower than expected. Governments spent heavily on catch-up programs - tutoring, summer school, technology, additional teachers. Some specific programs (high-dosage tutoring is the strongest candidate) have shown measurable effects, and the Education Scorecard, a Harvard and Stanford research project, credits federal relief money for most of the gains in the poorest districts since 2022. Even so, on the 2024 National Assessment of Educational Progress (NAEP), US fourth- and eighth-graders read 5 points worse than in 2019, and eighth-graders scored 8 points lower in math. By 2025, 9-year-olds' reading was back at its 2020 level; 13-year-olds' scores were not. Estimates of the lifetime economic effect on the most disrupted cohorts run into the trillions of dollars across affected countries, though they rest on assuming the shortfall is permanent.

What this exposes about the system. The pandemic showed both that in-person schooling is essential and that the system was less resilient than expected: the catch-up has been slower and harder than the recovery spending would predict. The system depends on assumptions - children show up, teachers teach, parents do not have to substitute for schools - that turn out to be more fragile than the pre-pandemic education conversation acknowledged. The first is still shaky: about 23% of US students were chronically absent in 2024-25, missing 10% or more of the school year, against about 15% before the pandemic.

The cost of college

US college tuition rose about 80% in real terms between 2000-01 and 2025-26 at public four-year universities and about 50% at private nonprofit ones (published "sticker" prices, College Board). Real median weekly earnings rose about 12% over the same period. College has become much more expensive relative to family income, and student debt has grown with it. The last decade looks different: published public four-year tuition has fallen 9% in real terms since 2020-21, and the average net price for first-year in-state students at public four-year colleges, after grants, fell from $4,450 in 2012-13 to about $2,300.

Several factors drive it, none of which alone explains the picture. State funding for public universities has fallen as a share of those universities' budgets, with tuition rising to fill the gap. Administrative spending has grown faster than instructional spending. The amenities arms race (better dorms, recreation centers, athletic facilities) has consumed significant capital spending. Federal student-loan availability has expanded faster than wages, which critics argue lets universities charge more without price discipline from buyers. Each of these is a partial cause; how much each contributes is disputed.

What this has produced. The federal government holds about $1.7 trillion of student loans, owed by 42 million people (June 2026). New York Fed credit-bureau data, which count differently, put all student debt at $1.65 trillion, up from about $240 billion in 2003. The median federal balance is around $20,000, but the long tail is severe: the 3.8 million borrowers who owe over $100,000 each, often law and medical school graduates, hold about 40% of the federal total. The debt has measurable effects on home ownership (a Federal Reserve study found that 10% more debt cuts home ownership by 1-2 percentage points in borrowers' first five years out of school), family formation, retirement saving, and career choice. Borrowers who do not finish have the worst outcomes - debt without the credential that justifies it - and nearly four in ten students who start college have no degree or certificate six years later.

The political response has swung. The Supreme Court struck down the Biden administration's broad loan forgiveness in 2023, and in 2025 a federal appeals court ruled its most generous repayment plan unlawful. A 2025 budget law then tightened federal lending. From July 2026 it ends uncapped federal loans for new graduate students (Grad PLUS), caps parent borrowing at $65,000 per student, and offers new borrowers only a standard plan or a single income-driven plan, with most older plans closing in 2028. The Congressional Budget Office expects its higher-education changes to save about $284 billion over ten years. Some states (Tennessee, New York, others) have made community college free or reduced cost.

K-12 stagnation and decline

Beyond the pandemic effects, K-12 outcomes in most rich countries were roughly flat for years and are now falling. The OECD's PISA tests measure 15-year-olds' ability in math, reading, and science; the 2025 round covered 91 countries and economies. It recorded the lowest OECD averages in PISA's history in math and reading: since 2015 reading has fallen 28 points and math 22, the equivalent of about a year and a half and a year of learning. The US was at the OECD average in math and above it in reading and science, though its sample missed some of PISA's standards, and it had the widest gap between strong and weak readers of any participating system. The East Asian systems (Singapore, Japan, Korea, Taiwan, Macao, parts of mainland China) kept their lead.

Why the stagnation. Education researchers have a reasonably consistent answer: the things that matter most for student outcomes are the quality of teaching, the support and structure children have at home, the peer environment in school, and the cumulative effect of many years of incremental learning. Most reforms either change the easy-to-measure surface (testing, accountability, curriculum standards) without improving teaching, or try to improve teaching without the political and cultural support to do it. The reforms that have produced measurable gains (high-dosage tutoring, sustained early-childhood programs like Perry Preschool, specific phonics-based reading instruction in some districts) have been hard to scale beyond their original settings. For the US decline since about 2013, the Education Scorecard's authors name two likely causes, looser test-based accountability and the rise of social media, but say it is unclear how much each contributed.

What is changing now. Phonics-based reading instruction (the "science of reading" movement) has spread across US states after decades of debate. Every state that improved in reading between 2022 and 2025 was implementing it, though many that adopted it have yet to improve. Phones are leaving classrooms: 36 US states ban or limit them. A Florida study found test scores rising in a ban's second year, after a first-year jump in suspensions; an English study found no link between school phone rules and students' well-being. The federal role is shrinking. The Education Department has cut its staff by nearly half since January 2025 and handed the management of programs, including aid for poor districts and special education, to other agencies. It says this gives "parents, students, and teachers more control over education"; states, districts and unions are suing, and in July 2025 the Supreme Court let the layoffs proceed while the case continues.

The future-of-work mismatch

The system that produces educated adults was largely designed in the late 19th and early 20th centuries to prepare workers for an industrial economy. Most of its structure has changed less than the economy it was meant to feed. That structure is 12 years of K-12 schooling, four years of college as the standard credential, and an academic curriculum that emphasizes abstract reasoning and writing.

Most jobs that pay well and do not require advanced degrees now require some form of post-secondary credential or specific technical skill. The traditional vocational pathways (apprenticeships, community college technical programs, employer-trained skill development) are weaker in the US than in most peer countries. The four-year degree that became the default route to the middle class after World War II still pays. In 2025 the median full-time worker aged 22 to 27 with a bachelor's degree earned $60,000, against $40,000 with only a high school diploma (New York Fed). But the degree costs far more than it did, and recent graduates' unemployment rate, 5.7% in mid-2026, has risen above the 4.1% for all workers.

Germany and Switzerland are the usual counterexamples. Their dual-track education systems (academic for some students, structured technical apprenticeships for others) are associated with lower youth unemployment, higher earnings for non-college-educated workers, and a stronger industrial base than US-style college-or-nothing systems. There is a catch: Eric Hanushek and colleagues found that the early job advantage of vocational training shrinks with age, most of all in apprenticeship countries. The political conditions that built and sustain those systems - employer involvement in education, cultural respect for technical work, cooperative labor-management traditions - are not easy to replicate in countries that lack them.

Now consider AI. Many of the traditional knowledge-work jobs that justified the four-year college credential are now exposed to AI substitution (covered on the AI and the Concentration of Power page). The professional skills that AI is least likely to substitute for - judgment under uncertainty, learning from limited data, physical and social skill, creative synthesis, and ethical reasoning - are not skills the typical college curriculum is currently designed to develop. Education systems that adapt to this may give their graduates better economic prospects than those that do not. Most systems are adapting slowly.

AI and what is changing in classrooms

AI has reached classrooms faster than the school system has adapted to it. In a Pew survey in late 2025, 54% of US teens said they had used AI chatbots for help with schoolwork, and 59% thought AI cheating happened at least somewhat often at their school. The first wave (2023-2024) was characterized by alarm about cheating; the second (2025 on) is starting to be about what AI changes in teaching and learning.

What AI does well in education. Producing first drafts and outlines that students can react to. Explaining concepts in multiple ways and at different levels of detail. Personalizing practice problems for individual students at their actual level. Acting as a patient one-on-one tutor for students whose schools cannot provide that. In a randomized trial in a Harvard physics course, students using a purpose-built AI tutor learned more, in less time, than in active-learning classes. The promise is real.

What AI does not do well in education. Building the foundational skills that come from struggle and persistence (writing one's own first draft, working through a math problem on paper, memorizing vocabulary). Replacing the social-emotional learning that happens in classrooms with peers. Substituting for the relationship between a teacher who knows a student well and that student. Identifying when a student is struggling for non-academic reasons. These are the things that good schools have always done and that no AI can fully replace. The risk is measurable: in a Turkish high-school trial, unrestricted GPT-4 access raised scores on practice problems but left students 17% worse on later exams without it; a tutor version with guardrails avoided the harm.

The integration question is not whether AI will be in classrooms - it already is - but how to integrate it without losing what schools do well. The most promising experiments combine AI as a tutor and practice tool with strong in-person teaching that focuses on the dimensions AI cannot substitute for. PISA 2025 points the same way: students who used AI to draft their writing scored lower, but frequent users who had been taught to judge AI output did better than those who had not. Most school systems are still figuring this out.

How country systems compare

Different countries' education systems perform very differently and reflect different political and cultural choices. The numbers below are rough.

The takeaway: different systems work for different reasons, and the relationship between spending and outcomes is loose. The countries long cited for improvement (Estonia, Poland, Portugal, parts of the UK) got there through specific focused reforms over many years rather than any single dramatic policy change, and gains can fade. The US case - high spending, average-to-good scores, severe equity gaps, world-class top tier - is one of the most distinctive in international comparison.

The paths from here

Dated September 2026, weeks after PISA 2025 recorded the lowest OECD averages in its history in math and reading. Education is one of those topics where policy moves slowly and consequences accumulate over decades. Each path below is one realistic shape the next 10-20 years could take.

The base case is slow drift in average outcomes, with localized improvements in specific places that have invested in serious reform. AI-aided learning is the most likely transformational variable. Vocational and apprenticeship pathways may expand modestly. College costs will keep producing political pressure; the 2025 loan caps are the first large test of tighter federal lending. The class gap may widen further before any countervailing force becomes politically possible.

Where serious analysts disagree

Education research is unusually robust in some areas (the effect of high-dosage tutoring, the importance of early-childhood development, the lower returns on specific kinds of school-spending increases) and unusually contested in others. Each reading below is held by named scholars worth engaging directly.

What the available evidence supports: the US system has problems that have produced flat or falling outcomes despite high spending, though researchers disagree about the causes. Small, intensive early-childhood programs show large long-run gains, while large public programs have a mixed record. Vocational and apprenticeship pathways ease young people into work, though the advantage may fade with age. AI may transform individual learning if integrated well, and unguided use can set it back. The credentialing question is real and is starting to be addressed by some employers.

What this means for you

Education touches every reader's life directly through their own schooling, their children's, and the workforce around them. A few practical observations: