El Salvador Gave A Million Kids An AI Math Tutor. And Then…

• El Salvador is deploying AI tutors across more than 5,000 public schools, the first national program of its kind.
• In a pilot across 171 schools, 1,198 students took the PISA for Schools test in June 2026 and scored comparably to Germany and Sweden’s 2022 results.
• El Salvador’s 2022 national PISA math score was 343. Germany’s was 475 and Sweden’s was 482.
• The caveats are real: the students volunteered, the sample is small, and this is not a randomized trial.
• Even discounted heavily, it’s the largest piece of evidence yet that one-to-one AI tutoring works outside a lab.

In 2022, El Salvador’s students scored 343 in math on PISA, the international assessment given to fifteen-year-olds. Germany scored 475. Sweden scored 482. About 63% of Salvadoran students performed poorly across reading, math and science.

Those are not gaps you close with a better textbook.

In June 2026, after just over a year of a pilot program, 1,198 students across 171 Salvadoran schools took PISA for Schools. They scored in the same range as Germany and Sweden had in 2022.

That’s the story, and it deserves both attention and scrutiny.

What they actually did

El Salvador announced a national AI education program in December 2025, with plans to reach more than 5,000 public schools over two years. Every student gets access to an AI tutor. Teachers get training and support tools.

The pilot ran first, in 171 schools, for roughly a year before the assessment. Then they tested.

The scale is the interesting part. Small AI tutoring studies have existed for years, run by universities on a few hundred kids under controlled conditions. Nobody had tried to hand an AI tutor to an entire country’s school system, and the usual assumption was that it would fall apart on contact with reality: bad connectivity, untrained teachers, kids who use it to cheat rather than learn.

What the numbers say

Pilot students scored above El Salvador’s national averages in reading, math and science, landing near Germany’s 2022 marks (475 math, 480 reading, 492 science) and Sweden’s (482 math, 487 reading, 494 science).

To put that in perspective: moving from 343 to somewhere near 475 is not an incremental improvement. On PISA’s scale, roughly 20 points is often described as about a year of schooling. A shift of that size in about a year of program time is larger than most education interventions produce in a decade.

Now the caveats, because they matter

The students volunteered. This is the big one. 1,198 volunteers out of 171 schools is a self-selected group, and the kids who sign up for an optional international test are not a random sample of Salvadoran fifteen-year-olds. Some of the gain is almost certainly selection.

PISA for Schools is not national PISA. It’s a related instrument designed for individual schools, and the results are not perfectly interchangeable with the country-level scores it’s being compared against.

There’s no control group. This is a before-and-after comparison against a national average from a different sample in a different year, not a randomized trial. A responsible researcher would call this promising, not proven.

A year is a year. Nothing here tells you whether the gains hold in year three.

Discount it as much as you think is fair. What’s left is still the largest real-world test of AI tutoring anyone has run, and it did not fail.

Why this shouldn’t be surprising

There’s a forty-year-old finding in education research called Bloom’s two sigma problem. In 1984, Benjamin Bloom reported that students taught one-to-one with mastery learning performed about two standard deviations better than students in a conventional classroom. The average tutored student outperformed 98% of the control group.

Bloom’s point wasn’t that tutoring works. Everyone knew that. His point was that it was economically impossible to give every child a personal tutor, and he challenged educators to find group methods that came close. Forty years of attempts produced improvements, but nothing at two sigma.

The Salvadoran program is the first serious attempt to solve Bloom’s problem by making one-to-one tutoring cheap instead of making group instruction better.

That reframe is worth sitting with. Every structural weakness of a classroom, thirty kids at different levels, one adult, a fixed pace set by the middle of the room, exists because personal attention was expensive. Change the cost and the constraint stops applying.

What it means for your kid

Careful here, because the leap from a national pilot to your fourth grader is longer than most people writing about this admit.

What transfers is the mechanism, not the headline. The reason one-to-one instruction beats a classroom is that it can stop, find out what a student doesn’t know, and teach that instead of what’s next on the schedule. That’s not exotic. It’s what a good tutor does, and it’s what your child’s teacher would do if they didn’t have twenty-nine other kids.

What also transfers is frequency. AI tutoring is available on Tuesday at 7pm and Sunday morning, which means a student can get four short sessions a week instead of one long one. That difference is most of what makes tutoring effective, and it’s usually priced out of reach when a human’s hours are the product. The cost comparison is stark.

What doesn’t transfer is any guarantee. A tool your kid doesn’t use does nothing, and no study changes that.

Where this is going

El Salvador will have several years of data on more than a million students. That will settle questions the pilot can’t, and some of the answers may be unwelcome.

For now it’s the best evidence available, and it points the same direction as Bloom did in 1984. One-to-one instruction works. The only thing that ever stopped it was the price.

That’s the bet Math Anon is built on. More about why we built it.

Related reading

• Why 20 Minutes a Day Beats a 90-Minute Tutoring Session
• How Much Does a Math Tutor Cost?
• How to Find the Gap: A Parent’s Guide to Math Diagnostic Assessment
• Is My Child Behind in Math? 12 Signs by Grade Level

Sources

• PISA for Schools results, June 2026, 171 schools, 1,198 students
• OECD PISA 2022 country results for El Salvador, Germany and Sweden
• Bloom, B.S. (1984), “The 2 Sigma Problem,” Educational Researcher