Opinion: Global test scores hit a two-decade low. Letting AI do students' thinking made it worse.

The Organisation for Economic Co-operation and Development released the largest and most discouraging round of results in the quarter-century history of its Programme for International Student Assessment this month, and the timing is not incidental. This was the first PISA cycle to test 15-year-olds after generative AI tools spread into homework routines worldwide, and the OECD's newly published report shows something more specific than a generic post-pandemic slump: the steepest score losses cluster precisely around the tasks chatbots now do for students — drafting a paragraph, summarizing a passage, solving an equation. That pattern is not proof of a single cause, and it should not be treated as one. But it is strong enough evidence that school systems, including America's, have let artificial intelligence into classrooms faster than they have figured out what it costs.

The argument here is not that phones and chatbots are solely responsible for a global education slide that has other, well-documented causes. It is that policymakers and school boards are treating the arrival of generative AI in daily schoolwork as a logistics problem — device policies, plagiarism detectors, acceptable-use forms — when the PISA data say it is also a cognitive one. Reading and mathematics are skills built through effortful repetition. When that effort is outsourced by default, the repetition stops happening, and the 2025 results are the first large-scale evidence of what that looks like at a population level.

A record slide, twenty years in the making

PISA 2025 tested roughly 760,000 students across 91 countries and economies, the largest sample in the exam's history. Average reading scores across OECD countries fell 28 points from 2015 to 2025, equivalent to about a year and a half of learning; math fell 22 points, or roughly a year. One in five 15-year-olds across the OECD is now a low performer across reading, math and science combined, up from 16 percent as recently as 2022.

The United States fits the broader pattern with its own specifics. American reading scores dropped 14 points since 2022 to their lowest level in more than two decades, according to the results the National Center for Education Statistics published this month; math fell 15 points since 2018, to an average of 463. Science held roughly flat. Meanwhile the wealthy Chinese provinces and municipalities of Beijing, Shanghai, Jiangsu and Zhejiang — the sample the OECD uses as its proxy for China, not a nationally representative one — posted the highest math average of any of the 91 participants, up 21 points since 2018. The trend lines are moving in opposite directions, and the gap between them now amounts to several years of learning. Not every country followed the same script: Turkey was the only OECD member to improve across all three subjects, evidence that decline is not preordained, and a breakdown of the exam's other findings shows Estonia holding near the top of the rankings throughout the slide.

The chatbot did the homework, and it shows

For the first time, PISA asked students directly how they use artificial intelligence for schoolwork, and the pattern that emerged is uncomfortable. Across the participating systems, more than half of students said they used chatbots at least several times a month for research, drafting or summarizing; in 82 of the education systems tested, fewer than half of students reported using AI "never or almost never" for schoolwork. Students who used AI daily to summarize what they read scored nearly 30 points below peers who rarely or never did, roughly a full year of learning. The exception mattered as much as the rule: frequent AI users who had been explicitly taught to evaluate and verify AI-generated output outperformed frequent users who had not, suggesting the damage comes from unsupervised offloading rather than the technology itself.

Andreas Schleicher, the OECD's director of education and skills and the architect of PISA, put the mechanism plainly in a recent interview, arguing that AI can substitute for the very struggle that produces learning.

"You don't get fit by watching sports. You get fit by doing sports. And the same is true with learning."

American data point the same direction. A national survey of high schoolers and recent graduates published this month by the Thomas B. Fordham Institute found 48 percent of respondents regularly let AI solve their math problems, more than a third used it to avoid reading assigned texts altogether, and roughly a quarter had used it during quizzes or tests; 58 percent agreed AI has made homework "meaningless." A review of how those habits map onto the exam results found the same warning signs playing out across the participating countries, not just in the United States.

The case against overreacting

The strongest objection to this argument is methodological, and it deserves to be taken seriously: correlation between AI use and lower scores is not proof that AI caused the decline. Students who are already struggling with a subject are more likely to reach for a chatbot to get through an assignment, which means the causal arrow could run backward — weak skills driving AI dependence rather than the reverse. Dan Goldhaber, an education researcher who studies these dynamics, has warned against assuming the technology itself is the culprit rather than a symptom of instruction and incentives that were already failing students. Screen time generally, chronic absenteeism, pandemic-era learning loss and teacher shortages all predate widespread chatbot use and plausibly explain a meaningful share of the slide on their own.

There is also a second, less flattering complication for anyone eager to declare an American crisis: the United States has actually gained ground in the international rankings over the past decade, but almost entirely because other countries fell faster. One widely read analysis of the rankings put it bluntly: American students now sit closer to the top of the pack, but only because peers elsewhere are sinking, not because U.S. classrooms are doing anything differently or better. That is a reason for humility on both sides of this argument — for those who want to treat AI as the singular villain, and for those who would use the improved American rank to wave the results away.

What comes next

The people most exposed to getting this wrong are not policymakers; they are 15-year-olds who will spend the next half-century in an economy that still rewards the ability to read a dense document, evaluate a claim and do arithmetic without a tool doing it for them. The OECD's own data show that guided, purposeful AI use can coexist with strong outcomes — it is unsupervised substitution, not the technology, that correlates with the losses. That distinction should drive policy rather than get lost in a broader panic about screens.

Concretely, that means school systems should sequence AI access the way they long ago sequenced calculators: withheld until foundational fluency is demonstrated, then introduced with explicit instruction in verifying and critiquing machine output, not left to arrive by default on a school-issued laptop in sixth grade. It means treating "AI literacy," which several countries now claim to teach, as distinct from mere AI access, and measuring the former rather than assuming the latter delivers it. And it means resisting the temptation, on display in some of the commentary around these results, to treat a shared global decline as proof that no one country needs to act. China's PISA sample gained a year's worth of math learning in the same years the United States lost most of one; whatever else explains that divergence, it did not happen by accident, and it will not be reversed by accident either. The PISA 2025 results are a warning delivered with unusual clarity. What happens next is a choice, not a foregone conclusion.