PISA 2025 Is a Bad Ranking. But It May Be a Very Good Alarm

Editorial image generated by the author.

PISA 2025 brought the OECD average to its lowest recorded level in reading, mathematics and science. The newest fall was concentrated in reading and mathematics: reading is down 28 points since 2015 and mathematics 22. One in five 15-year-olds across the OECD now performs below the baseline level in all three core subjects.

The obvious conclusion is that schools are failing. The awkward part is that PISA is not particularly good at telling us which countries have the best schools. National populations differ, immigration and socioeconomic background matter, participation rates vary, and students have little personal incentive to perform well on a test with no consequences for them. Even apparently precise country rankings contain enough uncertainty to make some movements up or down the table meaningless.

Yet those weaknesses do not make the long-term decline disappear. The important signal in PISA 2025 is less that one country has overtaken another than that many very different countries have deteriorated together. PISA may be telling us something important—just not quite what its rankings imply.

What PISA Measures — and Why the Rankings Mislead

PISA is often discussed as though it were an international examination of what children have learned at school. That is not really its purpose. Every three years, the OECD tests samples of 15-year-olds on their ability to apply reading, mathematical and scientific knowledge to problems. More than 760,000 students across 91 countries and economies participated in 2025. The aim is not primarily to test curriculum recall, but whether students can understand a text, extract relevant information, reason mathematically, evaluate evidence and apply what they know to unfamiliar situations.

These are useful abilities, but they are not the entirety of education. PISA tells us relatively little about historical knowledge, literature, artistic ability, vocational skills, civic understanding or many other things schools might reasonably be expected to develop. Nor can a national score cleanly isolate what schools themselves have achieved. A 15-year-old arrives at the assessment after fifteen years of family life, language exposure, preschool, school, friendships, books, television, gaming, social media and whatever else has shaped that child's development.

Sweden illustrates why population composition matters. According to Sweden's PISA 2022 country note, students with an immigrant background scored 81 points below non-immigrant students in reading and 63 points below them in mathematics. Socioeconomic differences explained part of those gaps, but not all of them.

That complicates international comparison. A system educating more recent arrivals, second-language learners or children whose previous schooling was disrupted can record a lower national average without every part of the school system having become worse. At the same time, those students remain part of the country's actual population. Adjusting them away may help answer a question about school effectiveness; it does not answer the different question of what skills the country's 15-year-olds actually possess.

Sampling creates another problem. New Zealand's 2025 participation fell sufficiently short of PISA standards that the OECD's non-response analysis estimates that its published mean scores may be overstated by roughly 10 to 15 points. A possible bias of that size should make anyone cautious about declaring an educational triumph because one country leads another by six points.

Questions must also work across languages and cultures, students differ in how seriously they treat a low-stakes assessment, and PISA uses statistical models to place students who answered different subsets of questions on a common scale. None of this makes the exercise useless. It does make the polished international ranking look more precise than the underlying measurement really is; the OECD itself reports confidence intervals for countries’ possible rank positions.

Why the Long-Term Decline Is Harder to Dismiss

The stronger criticism of PISA runs into a different problem. PISA is not literally the same examination repeated every three years, so long-term comparisons cannot simply be assumed to be valid. Frameworks evolve and assessment methods change.

Reading, however, gives the OECD an unusually good opportunity to test that objection. The 2025 reading assessment was based on the 2018 framework, and many of the same tasks had already been used in both previous cycles. The OECD's detailed analysis of the 2018–2025 reading trend examined task characteristics, adaptive testing, test position, motivation and response behaviour. Its conclusion is unusually direct: the decline is robust and unlikely to be a measurement artefact.

Across 35 OECD countries with comparable data, reading fell by 11 points between 2018 and 2022 and another 14 points between 2022 and 2025. Across all participating systems with comparable data, 56 of 74 recorded statistically significant declines over those seven years. Only five improved, and all five remained below the OECD average.

Changing immigration patterns do not make this wider trend disappear either. Across 29 OECD countries, the share of students with an immigrant background increased by 3.8 percentage points between 2018 and 2025, while the reading gap between immigrant and non-immigrant students remained around 40 points. Mechanically, that change in composition accounts for roughly 1.5 points of a 25-point fall in reading. Immigration can meaningfully affect individual countries and comparisons between them; it does not explain the scale of the OECD-wide deterioration.

The longer trajectories are also different enough to resist a single simple story. Reading began weakening before the pandemic. Mathematics remained comparatively stable for longer and then fell sharply after 2018. Science declined more gradually and was broadly stable between 2022 and 2025. Something wider than one failed national reform appears to be involved.

The Usual Explanations Do Not Fit Neatly

Explanation	What it plausibly explains	What it struggles to explain
Immigration	Part of national score differences and trends in some countries	Broad declines across countries with very different migration patterns
COVID disruption	A plausible contributor to the exceptional 2018–2022 deterioration	Declines that began before 2020 and continued afterwards
Generative AI	Possible weakening of some forms of independent work since 2022	A trend already established before widespread LLM use
Socioeconomic inequality	Large and persistent differences between students	The particularly large recent decline among advantaged students
National school policy	Differences between individual countries	Similar deterioration across many different systems
Changing digital and information habits	A common exposure across wealthy societies over roughly the relevant period	Causation remains unproven

COVID clearly belongs in the explanation. School closures and wider disruption coincided with the extraordinary fall recorded between 2018 and 2022. But reading performance had already been drifting downward before 2020 and then fell sharply again between 2022 and 2025.

Generative AI has an even clearer timing problem. Students can now ask an LLM to summarise a text, solve an equation, structure an essay or explain a scientific concept without necessarily performing the underlying intellectual operation themselves. That may matter considerably for what happens next. It cannot explain a deterioration visible years before ChatGPT existed.

Nor should the international pattern be used to absolve schools. Wealthy countries are not completely independent educational experiments. Many have experienced some combination of classroom digitisation, teacher shortages, changing curricula and assessment practices, rising absenteeism and shifts in parental engagement. A problem appearing across countries does not automatically mean its cause lies outside school. What the data do make difficult is the claim that one simple institutional failure explains the whole pattern.

The Most Troubling Clues Are Who Is Falling — and How They Read

One of the more surprising findings in the PISA 2025 student-performance data concerns the students who should have been best protected. Educational inequality remains large, but between 2022 and 2025 some of those gaps narrowed because advantaged students fell faster. In reading, disadvantaged students lost four points while advantaged students lost twenty. In mathematics, advantaged students lost fourteen points while the change among disadvantaged students was not statistically significant. In science, disadvantaged students were broadly stable while advantaged students fell by eight points.

That is not encouraging equality. It is compression through deterioration at the top. Children from affluent families are more likely to have educated parents, books, tutoring, extracurricular opportunities and well-resourced schools. If their performance is falling particularly quickly, explanations based mainly on lack of material resources become less convincing.

The reading test also records a change in how students answer. On one reading-fluency measure, the proportion classified as “hasty readers”—students giving fast and incorrect responses—rose from 6.6% in 2018 to 11.4% in 2025. Difficulties with longer texts remained a strong contributor to the overall decline. The fall was visible from the beginning of the assessment, so simple late-test fatigue cannot explain it, although losses were larger in some later sections.

One interpretation is that students simply care less about PISA. That would mean the assessment partly exaggerates the loss of underlying ability. Yet the OECD's deeper analysis does not find a general collapse in all measures of test engagement, and its conclusion is more specific: students have become less successful at tasks requiring careful reading and sustained attention.

That distinction matters because sustained attention is not merely an inconvenience introduced by the test. School, university and most professional work contain difficult texts and problems that offer no reward every few seconds. A student becoming less willing—or less able—to persist with such material may not be becoming less intelligent in any broad sense. The educational consequence can still be serious.

From Screen Time to Fragmented Attention

This is where the usual argument about “screen time” becomes too crude. An hour spent reading a novel on a tablet is not cognitively equivalent to an hour of short-form video, although both count as an hour looking at a screen. A teenager researching the causes of the First World War on a laptop is doing something very different from repeatedly switching between messages, videos and a game.

The more useful distinction may be between sustained engagement and fragmented consumption. A book asks the reader to remain with one argument or narrative for an extended period. A social feed continually offers another decision about whether the next item deserves a few seconds of attention. Search requires deciding what information is needed and locating it; recommendation systems increasingly decide what should appear next. Writing requires organising thoughts into language; autocomplete reduces some of that effort, while generative AI can sometimes remove much more of it.

None of these technologies is inherently harmful. Many are extraordinarily useful. The question is whether changing how often we practise certain cognitive activities changes how good we become at them. That is hardly an exotic proposition: abilities usually improve through use. It would be surprising if sustained reading, concentration and independent reasoning were completely exempt.

The chronology is certainly suggestive. The long deterioration in reading overlaps with the spread of smartphones, increasingly algorithmic social media and frictionless digital entertainment. Short-form video intensified the competition for attention; generative AI now allows parts of the reasoning, summarising and writing process to be outsourced as well.

Chronology, however, is not causation. Heavy social-media users differ from habitual book readers in many other ways, and PISA is observational evidence rather than an experiment. Family habits, personality, sleep, peers, prior academic ability and school quality all interact. The OECD itself is explicit that PISA cannot determine whether weaker reading results reflect the growing role of LLMs, video and other alternatives to long-form text.

Three claims therefore need to remain separate. Reading and mathematical performance have deteriorated across a large number of countries; the evidence for that is strong. Students also appear increasingly prone to rushed responses and difficulty with demanding, sustained reading tasks; PISA provides direct evidence for that. The claim that smartphones, algorithmic media or generative AI caused those changes remains a plausible hypothesis rather than an established result.

An Alarm, Not a Diagnosis

Suppose schools had remained exactly as effective as they were fifteen years ago, but children outside school gradually spent less time reading long texts, less time concentrating on one activity and less time solving informational problems without assistance. Academic performance could deteriorate even without a deterioration in teaching.

The reverse is also possible: schools may themselves have changed in ways that interact badly with the wider information environment. The point is not to move responsibility from schools to smartphones. It is that the boundary around “school quality” may be too narrow for the phenomenon PISA is now recording.

This is why the international ranking has become the least interesting part of the exercise. Whether Sweden should copy Estonia, Germany should emulate Singapore or Britain has moved several places up a table can produce useful policy debates, but PISA cannot support those comparisons with the precision often implied by the headlines.

The harder question is why teenagers across very different wealthy societies are increasingly struggling—or increasingly unwilling—to perform the same kinds of sustained reading and mathematical reasoning that earlier cohorts handled better. Immigration affects averages. COVID contributed to disruption. School systems matter. Test motivation matters. AI may shape what happens next. None of these explanations, on its own, fits the entire pattern.

The deeper possibility is that schools are trying to cultivate sustained concentration in an environment increasingly optimised to interrupt it, deep reading in an information culture organised more and more around fragments, and independent reasoning at a moment when answers are becoming almost frictionless to obtain.

PISA cannot diagnose that problem. Its own data cannot tell us whether the malfunction lies primarily inside schools, outside them, or in the interaction between the two.

But the uncertainty is increasingly about the diagnosis, not whether there is anything worth diagnosing.

PISA may be a poor ranking of school systems. It may nevertheless be a very good alarm.