While AI solves mathematics, teenagers are learning less

Erick Sergio Valenzuela AguilarCo-Founder & Head of Product

September 13, 2026 · 5 min read

CoverThirty out of a hundred: the Mexican 15-year-olds who reach basic proficiency in mathematics. Data: OECD, PISA 2025 (results released September 2026).

And this is probably the most uncomfortable of the three.

These are the most recent results available. The test was administered in 2025, hence the name PISA 2025, and the OECD published the results on 8 September 2026. They show a significant deterioration in the skills of 15-year-old students.

In reading and in mathematics, the OECD average is the lowest PISA has recorded since its first edition in 2000. Reading suffered the steepest fall: 28 points between 2015 and 2025.

In science, the main focus of this edition, the average held steady compared with 2022, after trending down since 2009.

The decline started before the pandemic. COVID-19 explains part of the problem, not the problem.

What the OECD points to

The report does not blame the decline on a single cause, but it does note that greater digitalisation, longer screen time and less reading for enjoyment have coincided with weaker reading results.

• Using digital devices to learn at school, in moderation, tends to go with better results. Using them for leisure at school goes with substantially lower results.
• On average, 28% of students say their classmates are distracted by digital devices in most or all science lessons.

There is one question PISA 2025 cannot yet answer with data: what happens when AI tools are used as a substitute for reasoning rather than as support. It is the conversation ahead, and it is better had before the answer shows up in the next edition.

A factor PISA does not measure: the class nobody reviews

There is a factor the report does not capture and that, from where we work, we consider crucial. With the pandemic, a large share of teaching moved to digital platforms: learning management systems (LMS), virtual classrooms and video classes. In many institutions that migration was not reversed when classrooms reopened: it stayed as part of day-to-day operations.

The problem is not the technology. It is that the class lost its witnesses. In a physical classroom, a coordinator can walk in, observe and correct. Hardly anyone rewatches a video class inside an LMS: the platform records that the session happened, not whether anything was taught in it. A school can pile up thousands of hours of recorded classes without a single piece of evidence about their quality.

Teaching went digital, but it did not become auditable.

It is worth being precise: PISA does not measure this factor, and the decline began before the pandemic, so it does not explain it on its own. But teaching that nobody reviews is fertile ground for the decline to deepen without anyone catching it in time.

And Mexico?

The news is not good either. In PISA 2025, Mexico scored:

• 388 points in mathematics (OECD average: 463).
• 408 in reading (OECD: 461).
• 414 in science (OECD: 482).

Compared with 2022, mathematics and reading each fell seven points and science rose four. Only the fall in mathematics is statistically significant, and at 388 points the country is back to a level similar to 2003.

The figure that weighs most is another one: only about 30% of Mexican students reach basic proficiency in mathematics, against around 65% across the OECD. In reading and in science roughly half reach it, against 69% and 74% in the OECD.

Mexico ranked 37th of the 38 OECD countries in reading and in science.

There are positive nuances. Science shows a slight recovery, and 66.1% of Mexican students say school prepared them for adult life, above the OECD average of 51%. But the structural problem remains.

The contrast is hard to ignore

At bottom, this week’s three stories are about the same thing. We are building tools that can reconstruct biological nervous systems, explore mathematics that has been open for almost a century and put thousands of artificial agents to work at the same time.

But our technological capacity is growing faster than some of our basic human capacities. We can ask an AI how to solve an extraordinarily complex equation, but we still need people who can read the explanation, question it and understand why it might be right.

Perhaps the great challenge of this decade is not building ever more capable artificial intelligence, but making sure our own capacity to learn, reason and understand grows alongside it.