PISA 2025 produced the worst results in twenty-five years of monitoring. For the first time, the assessment tested students who had grown up alongside chatbots, and the gap between those who hand their homework over to machines and those who do not amounted to an entire year of schooling.
On September 8, 2026, the Organisation for Economic Co-operation and Development released a figure that officials in Paris tried to present without drama: minus 28 points. That is how much the average reading proficiency of fifteen-year-olds across OECD countries fell over the decade from 2015 to 2025. By the organization’s own methodology, this is equivalent to roughly a year and a half of schooling erased from the educational lives of an entire generation. Mathematics lost 22 points. Science, the primary domain in this cycle, escaped with a decline of seven.
This is where things become genuinely interesting. More than 760,000 teenagers in 91 countries and economies took the assessment, the largest sample in the program’s history, representing approximately 33 million fifteen-year-olds worldwide. The children who sat down at computers in the spring and early summer of 2025 were born in 2009 and 2010. On November 30, 2022, when OpenAI made ChatGPT publicly available, they were thirteen. The previous cycle, PISA 2022, had assessed them before that date. What we are looking at is the first international measurement of a cohort that passed through adolescence in the presence of a machine prepared to think in its place.
The OECD phrased the result cautiously. I will put it more sharply: we now have the first statistically significant evidence that the uncontrolled introduction of generative artificial intelligence into schools is costing more than education ministries had assumed.
Twenty-Eight Points No One Will Give Back
The report’s arithmetic is stark. One in five fifteen-year-olds in OECD countries now falls into the low-performing category simultaneously in science, mathematics, and reading. Three years ago, the figure was 16 percent. An increase of four percentage points in a single cycle is not statistical noise. It is a collapse unfolding in slow motion.
Nearly half of all participants, 45 percent, fail to reach the baseline level of reading proficiency. This is not about the mechanics of reading. The fifteen-year-old described by PISA can recognize letters and form words. What that student cannot do is something else: hold a two-page argument in mind, distinguish a claim from a conclusion, or notice that a source contradicts itself. OECD Director for Education and Skills Andreas Schleicher called reading the foundation of all skills and said the sharp deterioration was a matter of serious concern. Over the past decade, he estimated, teenagers’ ability to work with complex texts has deteriorated by the equivalent of roughly two years of schooling.
Buried in the report is an indicator that escaped most news headlines and deserves to be brought into the open. The share of what might be called rushed reading, when a student moves through a text quickly but inaccurately, nearly doubled between 2018 and 2025, reaching nine percent. Nine percent of teenagers are physically unable to read slowly. They scroll through a page the same way they scroll through a feed.
Outside the OECD, the picture is less severe: reading fell by 16 points rather than 28, mathematics by eight rather than 22, while science remained stable. The explanation is simple and unpleasant. It is easier to fall from a greater height. Countries that started from a lower baseline and had lower smartphone penetration lost less because they had less to lose at the upper end of performance.
Cambodia and Rwanda posted the lowest results of the cycle, at roughly 300 points in each of the three subjects, compared with an OECD average of around 470 to 480. The gap between a fifteen-year-old in Cambodia and one in Singapore is no longer measured in years of schooling but in entire educational eras.
How Much Can We Trust This Yardstick?
Before drawing political conclusions from these figures, the instrument itself has to be assessed. I will do that candidly, because PISA has serious substantive critics.
In May 2014, a group of academics led by Heinz-Dieter Meyer of the State University of New York published an open letter to Andreas Schleicher accusing the program of encouraging short-term thinking among education ministries through its three-year testing cycle and of narrowing education to what can be measured. Danish statistician Svend Kreiner separately challenged the validity of the scaling model used to rank countries. The organization has repeatedly encountered sampling problems: Spain’s 2018 reading results had to be withheld because of anomalous response patterns, while China’s participation is limited to four of its most developed jurisdictions, making direct comparisons between China and entire countries methodologically misleading.
My conclusion about the instrument is this. PISA is mediocre as a tool for producing precise national rankings, and a difference of two or three positions between Germany and Poland means almost nothing. But for measuring change within the same system over a decade, it is the best instrument available because the methodology and age cohort remain constant. A 28-point decline in the same population over ten years means exactly what it appears to mean, regardless of disputes over rankings.
The First ChatGPT Generation Took the Test, and the Result Is Not Encouraging
Here is the figure that alone justified conducting the entire cycle: 46 percent of students in OECD countries use chatbots for schoolwork at least once a week. Nearly half the classroom. No educational technology in history, neither television, nor the calculator, nor the interactive whiteboard, all of which were once surrounded by enormous expectations, achieved that level of penetration in three years.
The OECD then did something education ministries had not done: it separated different types of use.
Students who use artificial intelligence for specific academic tasks, such as identifying the main idea in a text, writing an assignment, or conducting research, performed worse in science than students who do not use artificial intelligence. The gap is equivalent to approximately one year of schooling. One year. In exchange for homework submitted on time and, quite possibly, graded well.
Using artificial intelligence for broader purposes does not produce the same decline and in some cases is associated with a modest improvement in science performance. The distinction is fundamental, and it needs to be stated plainly because future regulatory policy depends on it. A machine that explains can help. A machine that performs the task can harm. The dividing line is not the technology itself but who is actually doing the cognitive work.
The term the report’s authors use in their explanatory section is cognitive offloading. A student systematically transfers the search for and processing of information to technology instead of doing it independently. The effect has long been known in neuropsychology and has been observed in far more innocent settings: people who photograph museum exhibits remember them less well than those who simply look at them. The brain keeps accounts. What it does not need to retain, it does not retain.
Add to this the 28 percent of teenagers in developed countries who reported that classmates are constantly distracted during science lessons by checking their devices. A physics lesson in which a quarter of the class is on a phone is not a physics lesson in any meaningful sense.
A Vaccine No One Tested
The metaphor the OECD allowed itself to use in the report is unusually harsh for an organization of its kind. The mass introduction of artificial intelligence into schools is compared with administering a vaccine on a mass scale before it has undergone the necessary trials.
This should be read as an indictment of the agencies responsible, not merely as a scientific observation. The authors place responsibility for the speed of adoption on institutions that did not have, and still do not have, sufficient data on how artificial intelligence affects children. The wording is cautious. The meaning is not: education ministries allowed a tool into classrooms whose effects no one had measured, and three years later discovered the consequences through external testing.
The pharmaceutical comparison works better than it first appears. No drug reaches the European market without three phases of clinical trials followed by post-market safety monitoring. Generative models entered schools without a single phase. Regulation was structurally absent: the European Union Artificial Intelligence Act, adopted in 2024, classifies educational artificial intelligence systems as high-risk, but its provisions are being phased in and primarily address systems used for assessment and selection, not the teenager who asks a model at midnight to summarize a textbook section.
Schleicher called on governments to establish clear rules for the use of artificial intelligence in schools and to determine where technology assists the teacher and where it becomes a shortcut allowing students to obtain ready-made answers more quickly. The appeal is correct and exactly three years late.
The economics behind that delay are easy to understand. Introducing digital platforms into schools is politically attractive: it is inexpensive compared with raising teacher salaries, provides an instant image of modernization, and allows governments to report visible progress to voters without undertaking structural reform. Education International, the global federation of education unions, cited chronic teacher shortages among the factors behind the decline. A tablet in a child’s hands costs less than keeping a physics teacher from leaving for the private sector.
The British Anomaly That Breaks the Convenient Narrative
At this point I am obliged to introduce a fact that works against simplistic conclusions. The United Kingdom ranked among the top ten education systems overall and improved its science results compared with 2022. At the same time, a relatively large share of British teenagers said they had never used artificial intelligence for school assignments at all.
The correlation is tempting, which is precisely why it cannot be presented as proof of causation. British schooling differs from continental systems in far more than its policy toward digital devices. England, whose education system differs substantially from those elsewhere in the United Kingdom, consistently outperforms Scotland, Wales, and Northern Ireland in every subject. That internal gap has persisted across multiple cycles and is associated with curriculum design, the Ofsted inspection system, and the systematic phonics approach to reading instruction introduced in English primary schools after 2006. PISA data cannot isolate the effect of limited artificial intelligence use from the effect of those reforms.
The cautious conclusion is this: the British case does not prove that restricting artificial intelligence improves educational outcomes, but it does rule out the claim that lagging behind in digitalization automatically means lagging behind in educational quality. For ministries that have spent the past five years measuring progress by the number of tablets purchased, that is an uncomfortable conclusion.
Top Students Are Drowning First
The most politically explosive finding in the report received almost no media attention. Socioeconomic inequality in educational outcomes across OECD countries narrowed between 2022 and 2025.
It sounds like good news until one examines the mechanism. The gap narrowed because children from affluent families performed worse, not because children from poorer families improved.
Equalization downward. That is what erosion looks like, not progress.
The logic is clear. A child from a low-income family has always studied in an environment with limited access to tutors, books, and parental attention, and that child’s performance has long hovered near the lower boundary. A teenager from a wealthy household possessed the very resource that smartphones and chatbots have devalued fastest: time for sustained reading, quiet, and disciplined attention. Technology struck privilege harder than deprivation because privilege consisted largely of time and concentration.
The United States illustrates this vividly. American students lost 14 points in reading compared with 2022, while 25.7 percent were classified as low performers and 13.3 percent as high performers. A country with one of the world’s most expensive school systems ranks thirteenth overall and twenty-seventh in mathematics, trailing even Germany.
Why East Asia Is Not Falling
The top five performers in the cycle have not changed and are unlikely to change in the foreseeable future. The highest results across the three subjects were recorded by the Chinese jurisdictions of Beijing, Shanghai, Jiangsu, and Zhejiang, together with Singapore. Estonia, Japan, South Korea, Macao, Taiwan, and the United Kingdom also ranked in the top ten.
European commentators usually explain this through examination culture and parental pressure. That explanation is only half correct. The other half is administrative, and it matters far more to anyone hoping to replicate the success.
In Singapore, teachers are recruited from the top third of university graduates, while the National Institute of Education is the country’s sole teacher-training pathway, giving the government complete control over the content of teacher education. East Asian systems regulate smartphone use in schools more strictly than European systems do, while extracurricular tutoring remains a major household expenditure rather than an act of charity. The Chinese jurisdictions participating in PISA represent the country’s wealthiest and most urbanized provinces, which systematically inflates China’s results relative to its national average. The methodology itself openly warns about this, and commentators routinely forget it.
Turkey, the United Arab Emirates, Montenegro, Qatar, and Slovakia all improved compared with previous cycles. Costa Rica posted a notable improvement in science. These systems have one thing in common: they are still in a catch-up phase, where even basic improvements in governance can produce rapid gains. It is easier to rise from 400 to 430 than to remain at 520.
Germany’s Trauma and Europe’s Fatigue
Germany ranked twenty-fourth, only slightly above the OECD average, and fell to twenty-sixth in mathematics. For a country that experienced a national shock over PISA twenty-five years ago, this is a diagnosis of the system, not an unfortunate fluctuation.
In December 2001, when the first PISA 2000 results were published, German fifteen-year-olds scored below the OECD average, giving rise to the term “PISA shock.” This was followed by educational standards introduced through the Standing Conference of the Ministers of Education and Cultural Affairs of the German states, national performance requirements, expanded all-day schooling, and programs supporting children from immigrant backgrounds. The reforms produced results in the mid-2000s and lost momentum by the mid-2010s. Germany is now returning to where it started, even though education spending has risen many times over in absolute terms.
Among OECD countries, the steepest declines in reading were recorded in Slovenia and Iceland. Iceland is particularly instructive: a country with near-universal internet penetration, one of the highest levels of school digitalization in the world, and a population comparable to that of a single district of Baku. Small size and wealth do not protect against the erosion of reading ability.
Europe’s problem runs deeper than methodology. For a decade, the continent invested in digital school infrastructure on the assumption that access to information was equivalent to the ability to process it. PISA 2025 demonstrated that these are different things, and that the latter deteriorates precisely as the former expands.
Ukraine: War Without Collapse
Ukraine’s case deserves separate examination, and the reason extends beyond the war.
Ukraine’s average score was 448 in science, 431 in mathematics, and 418 in reading. That puts the country 34, 32, and 43 points below the OECD average, respectively. Its combined score of 1,319 across the three subjects places Ukraine above Bulgaria, Moldova, and Georgia but below Slovakia, Poland, and Estonia.
The internal figures are more alarming than the averages. Forty-six percent of Ukrainian students fail to reach the baseline level in mathematics, 45 percent in reading, and 35 percent in science. At least one low score is recorded among 56.4 percent of students. In all three subjects simultaneously, the figure is 29.3 percent, nearly one in three.
Now for what is genuinely positive in these numbers, without stretching the facts for propaganda purposes. After the 2018 to 2022 decline, results stabilized: there was no further significant deterioration in 2025. Ukraine participated in the assessment for the third time and, for the second time, under conditions of full-scale war. An education system that lost territory, part of its teaching workforce, and millions of students to emigration or bomb shelters managed to hold its performance at roughly the level recorded three years earlier. From the standpoint of institutional resilience, that is an outcome worthy of study independently of the political context.
One line in Ukraine’s national report reads almost ominously: 54 percent of Ukrainian teenagers use artificial intelligence for schoolwork at least once a week, more frequently than their peers in advanced economies. The explanation is obvious. Remote learning, air-raid alerts, teacher shortages, destroyed schools: the chatbot substitutes both for homework and for the teacher who is physically absent from the screen. A country in which 29.3 percent of teenagers fail to reach baseline proficiency in all three subjects simultaneously also has one of Europe’s highest rates of artificial intelligence use in education. Whether this is coincidence or causation cannot be established from a single testing cycle. The hypothesis can be tested in 2028.
Who Will Pay for an Undereducated Generation?
Discussions about PISA are usually framed as a matter of educational anxiety. They should instead be framed as a matter of macroeconomics.
Education economists Eric Hanushek and Ludger Woessmann demonstrated through cross-country data in the 2010s that measurable cognitive skills, rather than formal years of schooling, explain long-term differences in gross domestic product growth rates. The logic of their calculations is straightforward: a country that raises its average PISA score by one-quarter of a standard deviation gains a meaningful increase in economic growth over several decades. The reverse side of the same model describes what is happening now. Losing the equivalent of a year and a half of schooling across an entire cohort is a deferred deduction from productivity that will begin to appear in economic statistics around the mid-2040s, when today’s fifteen-year-olds form the core of the workforce.
The developing world pays twice. According to World Bank estimates, roughly 70 percent of ten-year-old children in low- and middle-income countries cannot read and understand a simple text. In Washington, this indicator is called learning poverty, and it worsened sharply after the pandemic. The demographic dividend in Africa and South Asia, on which growth forecasts through 2050 have been built, works only if the young population is capable of learning. Cambodia and Rwanda, with scores around 300, show what that dividend becomes when this condition is absent.
There is also a geopolitical dimension that usually disappears behind educational rhetoric. The race for artificial intelligence is discussed in terms of chips, data centers, and energy. All three resources can be purchased. A fourth cannot: an engineer capable of reading a forty-page technical specification and identifying a contradiction in it. A country whose fifteen-year-olds are losing the ability to engage in sustained reading will discover a shortage of precisely that kind of professional fifteen years from now, at exactly the moment when it is trying to build a sovereign technological base.
Three Scenarios Through the End of the Decade
The first scenario is regulatory. Governments follow the pharmaceutical analogy to its logical conclusion: they restrict smartphones in schools, require mandatory labeling of artificial intelligence tools intended for minors, and demand evidence of educational benefit from developers. Precedents already exist. France banned mobile phones in schools for students under fifteen through legislation enacted on August 3, 2018. In its 2023 Global Education Monitoring Report, UNESCO explicitly recommended banning smartphones in schools. Australia imposed statutory restrictions on social media access for children under sixteen. Brazil introduced restrictions on phone use in schools in 2025. The direction has been established. The pace is inadequate.
The second scenario is inertia, and it is the most likely. Ministries issue framework recommendations without enforcement mechanisms, schools resolve the question at the discretion of individual principals, and the educational application market continues growing at double-digit annual rates. By the 2028 cycle, OECD countries lose another ten to fifteen points in reading, and we become accustomed to the idea that this is simply how things will be from now on.
The third scenario is adaptive, and it requires schools to do something they currently do poorly: redesign assessment. As long as homework remains the primary form of evaluation, chatbots will do it, because a rational teenager will always choose the cheaper path. A school in which written assignments are completed by hand in the classroom while books are read at home makes artificial intelligence substitution worthless without imposing a single ban. Such restructuring is expensive in terms of staff time and produces no attractive modernization statistics, which is why it is unlikely to be adopted on a mass scale.
Here are my predictions, each of which can be checked against specific dates. By April 2027, when the OECD publishes the results of its English-language proficiency assessment, followed in May 2027 by data from the Learning in the Digital World module, the same divergence will be confirmed: teenagers who use artificial intelligence as an explanatory tool will outperform those who use it to perform tasks for them. Formal rules governing the use of generative artificial intelligence in schools will be in place in most OECD countries by the end of 2028, but fewer than one-third will have mechanisms to verify compliance. The average reading score in PISA 2028 will not return to its 2018 level in any Western European country.
What This Means for Our Region
The South Caucasus occupies an uncomfortable position in this story. Georgia’s combined results place it below Ukraine, a country in its fourth year of war. Moldova and Bulgaria are in the same range. A region that has spent the past two decades developing strategies for transitioning from resource-based economies to knowledge economies is discovering that the basic condition for such a transition, a teenager capable of reading a complex text, is being developed less effectively than strategic documents had assumed.
Azerbaijan participated in previous cycles of the program, and its national results available in the public domain have been discussed extensively by the professional community. The question facing Baku is no different from the one facing Berlin or Tallinn, and it cannot be solved by purchasing equipment. Age restrictions for social media, which have been discussed in Azerbaijan’s parliament, belong to the same logic of protecting attention as France’s 2018 law. The logic is sound. Implementation determines everything.
I grew up in a country where reading a thick book was not a pedagogical exercise but simply a way to spend an evening because there were few alternatives. Long-term PISA data confirm what once seemed self-evident: the strongest results are consistently recorded among teenagers who read substantial books. This kind of reading trains the ability to retain and process large quantities of complex information. We gained access to all of humanity’s information while simultaneously losing the skill without which that access is useless.
The Conclusion Education Ministries Will Not Like
The OECD report is written in the language of international bureaucracy, which means its central conclusion has to be stated on the authors’ behalf.
The problem is not artificial intelligence. The school as an institution was built around information scarcity, while the world has moved into an age of information abundance, and not a single education ministry has rewritten its rules for that reality. Artificial intelligence did not create the crisis. It exposed a structure that had depended on ready-made answers being difficult to obtain. Once obtaining an answer became easy, it became clear that the effort we called learning had often been a byproduct of the difficulty of finding the answer rather than the result of deliberate educational design.
An uncomfortable conclusion follows. No ban on chatbots will restore anything by itself, because what has to be restored is not discipline but the purpose of the assignment. A fifteen-year-old who is given a two-page text and told to summarize it is entirely justified in considering the task solved when a machine does it in four seconds. The student is right. The assignment really is obsolete.
Three years from now, in 2029 or 2030, the OECD will publish the next cycle. The children who will take it are now in sixth grade and already have no memory of a world without generative models. Education ministries have roughly thirty-six months to decide what exactly they intend to measure in a generation for whom obtaining an answer is no longer an intellectual operation.
For now, the same scene is playing out in thousands of classrooms from Reykjavik to Kharkiv. A teacher hands out a text. Twenty-eight percent of the class lowers its gaze toward the desk, where a phone is hidden underneath. Nine percent scrolls down the page without finishing the paragraph. The lesson proceeds according to schedule.