A Decade of Decline
A week ago, in this same space, we wrote about a complaint shared almost unanimously by university professors: students are arriving in their first year of university increasingly underprepared. Their reading and mathematics skills simply are not up to the task. Then, along came PISA to confirm it, and we used some of its findings in that article to add context to the discussion. In this post, we take a closer look at the OECD report. What do the data tell us? What lies behind the figures? What is happening with screens and artificial intelligence? And what can we learn from the education systems that are still moving forward?
There is a very simple way to tell the story of PISA 2025. In mathematics, students across OECD countries score 22 points lower today than they did a decade ago. In reading, they score 28 points lower. The OECD translates those differences into more tangible terms: just over a year of learning lost in mathematics and around a year and a half in reading. One in five 15-year-olds now also performs poorly in Science, Mathematics and Reading simultaneously. In 2022, the figure was 16%.
But these numbers tell only part of the story.
PISA 2025 arrives at a particular moment. Five years have passed since schools closed because of COVID-19, and three since PISA recorded the steepest decline in performance in its history. At the time, the pandemic accounted for much of the explanation. Millions of students had lost months of face-to-face schooling, education systems had improvised remote-learning arrangements, and the consequences had been deeply uneven.
In 2025, we are still waiting for the rebound. Science, the main domain assessed in this edition, remains broadly stable compared with 2022 across the OECD. Mathematics has lost another nine points and reading another fourteen. In both areas, average performance has fallen to the lowest levels ever recorded by PISA.
Taking the longer view also changes the diagnosis. Much of the decline had already begun before 2020. The pandemic accelerated an existing trend and exposed problems with deeper roots: difficulties in securing foundational learning, teacher shortages across many education systems, changing reading habits, persistent inequalities and a rapid transformation in the way teenagers access information and engage with knowledge.
Stop here for a moment. You have just read roughly 400 words. If you were a 15-year-old, from this point on you might begin to struggle to understand the text, sustain your attention, connect ideas and reconstruct the meaning of what you have read. And we are only at the introduction.
So, although PISA 2025 deals in scores, it invites us to ask a much broader question: what is happening to learning among a generation that studies, reads, searches for information and solves problems in ways that are profoundly different from those of just ten years ago?
Reading in the Age of Scrolling
Of the three major competencies assessed by PISA, reading has lost the most ground. And that matters because reading is not simply one competency among many: it is the gateway to almost all the others. Reading is necessary to understand a mathematics problem, follow a scientific explanation, compare two arguments or decide whether a source is reliable.
The data allow us to go further. Between 2018 and 2025, the proportion of readers able to respond both accurately and fluently fell by seven percentage points. At the same time, hasty readers —students who move quickly through a text and answer fast, but incorrectly— rose from around 7% to 11%. Performance also deteriorates as students progress through the test. The OECD interprets both patterns as signs of a reduced ability to sustain attention, persevere or maintain confidence when faced with difficult tasks.
And this is where we need to qualify those 400 words we have just read. PISA does not say that a teenager suddenly stops understanding at word 401. What it does find is a greater decline when students face longer texts and, above all, in the skills those texts tend to require: evaluating information, making connections across different sources, interpreting and reflecting on what they have read. Across the OECD, barely 6% reach the highest levels of reading proficiency, where students are expected to understand lengthy texts, handle abstract or counterintuitive concepts and distinguish fact from opinion using implicit cues.
These skills are particularly important in an environment where information increasingly reaches us in fragments: messages, videos, headlines, posts and hyperlinks. The OECD observes that declining rates of reading for pleasure and increasing screen time have coincided with poorer performance, although PISA does not establish a causal relationship between the two.
Never before have we had so much information at our fingertips, or so many tools to find, summarise and organise it. But the easier it becomes to obtain an answer, the more important it becomes to know what to do with it: understand it, connect it with what we already know and question it when necessary. In a world of instant answers, understanding still takes time.
The Gap Is Narrowing… From the Wrong End
The gap between socioeconomically advantaged and disadvantaged students has narrowed.
At last, some good news… or is it?
In Science, advantaged students still outperform disadvantaged students by 85 points on average across the OECD. The gap has narrowed since 2022. That much is indisputable. But should we celebrate? No, because the gap has narrowed more because the students at the top have performed worse than because those at the bottom have improved. The same pattern can be seen in Mathematics and Reading.
Understanding this requires exactly the kind of operation PISA attempts to measure: connecting several pieces of information and resisting the most immediate interpretation. A smaller gap between two groups can mean that those at the bottom have moved up; it can also mean that those at the top have moved down. The figure is the same. The story is radically different.
Socioeconomic background remains one of the strongest predictors of performance. But its influence varies considerably from one education system to another. PISA identifies a small group —Canada, Estonia, Ireland, Japan, Korea, Macao and the Chinese regions of Beijing, Shanghai, Jiangsu and Zhejiang— that achieves something particularly difficult: strong performance combined with a relatively weaker influence of family background. In all of them, more than 75% of students reach at least the basic proficiency level in Science, while socioeconomic status explains less than 10% of the variation in performance.
How do they do it? PISA does not allow us to establish a common recipe or attribute these results to any single policy, but it does provide some clues. Several of these systems have classrooms with fewer disruptions and distractions, stronger teacher support and conditions that make it easier to identify and address learning difficulties. Japan, Korea and the participating Chinese regions, for example, combine highly disciplined classroom environments with very low levels of digital distraction. These systems differ greatly from one another, so searching for a single formula matters less than recognising the outcome: social disadvantage matters, but how much it matters also depends on what happens inside and around the school.
This is where academic resilience comes in. PISA uses the term for students who belong to the most socioeconomically disadvantaged 25% in their country yet still manage to rank among the top-performing 25%. They provide statistical evidence of an important distinction: inequality at the starting line and inequality in outcomes are not the same thing. Schools cannot erase the circumstances into which a child is born, but some education systems manage to make those circumstances matter considerably less in determining how far that child can go.
PISA 2025 confirms that the decline in learning did not start with the pandemic, nor dis it end with it.
Technology Is Not the Question. How We Use It Is
Any attempt to explain how teenagers learn today inevitably leads us to technology. Here, PISA 2025 introduces an important nuance: its data support neither the idea that more technology automatically produces more learning nor the claim that screens alone are responsible for the decline.
Moderate use of digital devices for learning at school is associated with better results. The relationship changes when devices are used for leisure during lessons. Twenty-eight percent of OECD students say their classmates are distracted by digital devices in most or all Science lessons. Students in those environments score, on average, 11 points lower in Science, even after accounting for the socioeconomic profile of students and schools. In Spain, where 29% report this kind of distraction, the difference reaches 16 points.
The data therefore shift the focus: what matters is not so much how much technology is in the classroom, but how it is used and for what purpose.
Artificial intelligence takes this much further. Forty-six percent of OECD students already use AI chatbots at least once a week to help them learn. Students who say they use AI to write assignments perform worse than those who do not. But among frequent users, another revealing finding emerges: students who have received training in how to evaluate the quality of AI-generated information tend to perform better than those who use these tools without such guidance.
The difference is small, but it introduces an important nuance: artificial intelligence is already part of the way many students learn. The challenge is to teach them how to use it without outsourcing to the technology the very things they need to learn to do for themselves.
PISA has begun exploring precisely this territory with its new Learning in the Digital World assessment. And it is worth clarifying what PISA means here by digital competence. The assessment does not measure whether a teenager knows how to operate a device. It asks whether students can use computational tools to learn something they do not yet know and solve a problem that does not have an immediate answer.
To do so, students work in digital environments where they are given tutorials, examples and feedback. They have to experiment, break a complex problem into smaller steps, build models or solutions, test whether they work, identify errors and correct them. In other words, they have to do something much closer to learning than simply following instructions.
The first results reveal a significant gap. Sixty-four percent of OECD students reach the level considered proficient in computational problem-solving. But only around half manage to combine that ability with basic proficiency in Science, Reading and Mathematics. Put another way, some students are reasonably capable of navigating computational tools while still lacking some of the foundational knowledge on which that use should be built.
The distinction is particularly relevant as artificial intelligence rapidly expands. Knowing how to use a tool and knowing how to think with it are not the same thing. To build a model, you need to understand what it represents; to interpret the data produced by an application, you need to know what those data mean; to spot an incorrect answer from an AI system, you need knowledge against which to check it. Digital competence, then, does not replace Reading, Mathematics or Science. It depends on them and, when used well, can take them further.
The Ones Moving Forward
The average is falling, but not everyone is falling with it. Türkiye, for example, improved in Science, Mathematics and Reading between 2022 and 2025. Looking back to 2015, the proportion of low-performing students there has fallen by 16 percentage points in Mathematics, another 16 in Reading and 25 in Science. Qatar records some of its best-ever results in 2025 and scores significantly higher in all three subjects than it did a decade ago. Montenegro has also improved in all three.
There is no formula that can simply be transferred from one country to another. Nor does PISA allow us to attribute an improvement to a particular policy. But the OECD does point to elements that recur in systems with stronger trajectories: focusing the curriculum so that content can be explored in greater depth, supporting teachers, directing resources towards those who need them most, involving families and using technology with clear pedagogical goals.
These are hardly spectacular conclusions in an age fascinated by disruption. True. But they are difficult to argue with. PISA 2025 arrives just as artificial intelligence promises to transform the way we search for information, write, calculate and solve problems. Its findings remind us that taking advantage of that transformation requires strengthening foundational skills rather than taking them for granted. To ask a machine to summarise a text, you need to know how to judge whether it has understood it. To accept a calculation, you need to understand what it means. To use an answer, you need knowledge against which to test it.
Ten years ago, a teenager had to search for many of the answers they can now obtain in seconds. In another ten years, they will probably be able to delegate many more intellectual tasks. The big question for education will be deciding which ones are worth delegating and which ones we must continue to teach precisely because the effort involved in doing them is part of learning.
PISA 2025 therefore leaves us with one undeniable truth and one possibility. The reality is that we have spent a decade losing ground. The possibility lies in the education systems that are managing to move forward while the average falls, and in technology that, when well integrated, can expand opportunities to learn. The task now is to understand what those systems are doing and how to use that technology so that, three years from now, the rebound is no longer something we are still waiting for.
