Countries Are Racing to Replace Teachers With AI Tutors. The Evidence Isn't Ready.

El Salvador just handed its entire public education system to Elon Musk's AI company. One million students. 5,000 schools. Grok as their personal tutor. Nobody knows if it'll work.

They're not alone. OpenAI launched "Education for Countries" at Davos in January — ChatGPT Edu deployed to eight nations including Kazakhstan (165,000 educators), Estonia, Jordan, Greece, and the UAE. At least ten countries are now betting their classrooms on AI tutoring with almost no track record.

The global education experiment just started. The control group doesn't exist.

One Study. 194 Students. A Lot of Weight.

The case for AI tutoring rests on a single paper. Harvard ran a randomized trial in 2025 — AI tutor vs. active learning in undergraduate physics. The AI group learned more in less time. They felt more engaged too.

Published in Scientific Reports (Nature), it got enormous attention. "AI tutoring outperforms in-class active learning" — the headline practically wrote policy.

But here's what gets lost. 194 students. At Harvard. In physics. Already high-performing, digitally fluent, self-motivated. Extrapolating that to a million kids in El Salvador's underfunded schools is a leap the researchers themselves didn't make.

What the OECD Actually Found

The OECD's Digital Education Outlook 2026 is more careful. GenAI can improve learning — but only "when guided by clear teaching principles." Without structure, students just outsource their thinking to the AI.

Two numbers stand out. English secondary science teachers using AI cut lesson planning time by 31%. And low-experience tutors using AI support saw a 9-percentage-point increase in student pass rates.

Both findings come with a caveat the size of a continent: they measured AI assisting teachers, not replacing them. The countries deploying national programs are doing something different. They're giving AI directly to students — often in classrooms where teachers are already stretched thin.

The El Salvador Gamble

El Salvador's education system was struggling before AI showed up. Outdated infrastructure, overcrowded classrooms, underpaid teachers burning out. The country spends less per student than almost any Latin American neighbor.

Into that gap steps xAI. Bukele announced the partnership in December 2025. Grok will "adjust to each student's pace, preferences, and mastery level." The pitch sounds like a Silicon Valley deck: personalized learning at scale, teachers as "collaborative partners," AI doing the heavy lifting.

What's missing: any pilot data. No trial in Salvadoran schools. No test of whether Grok works in Spanish for kids reading below grade level. No connectivity assessment across 5,000 schools. Deployment timeline: two years. Evidence timeline: hasn't started.

Kazakhstan, Estonia, and the OpenAI Cohort

OpenAI's approach is more structured but equally untested. "Education for Countries" gives governments ChatGPT Edu and GPT-5.2 classroom tools. Kazakhstan alone is training 165,000 educators.

The first cohort spans wildly different contexts. Estonia (one of Europe's top-performing systems). Jordan (the world's second-largest refugee population per capita). Trinidad and Tobago (200,000 students). The UAE ($2.7 billion AI strategy).

What works in Tallinn might not work in Amman. The OECD's warning was blunt: "Pushing these tools into under-resourced systems without solving basics such as devices, bandwidth, and teacher time is poor policy and risks deepening existing divides."

The Real Question Nobody's Asking

Here's what should bother you. These aren't app downloads. These are national education systems — the infrastructure that shapes how millions of kids learn to think, read, argue, and solve problems. Countries are signing deals before a single large-scale study has shown this works in K-12.

Eedi, a UK tutoring platform, ran a 2025 study showing AI-plus-human tutors matched human-only tutors. Matched. Not beat. They're planning a bigger trial for 2026 — meaning even the companies building these tools don't have proof yet.

The gap between "this might work" and "we're deploying it nationwide" is enormous.

What's Actually at Stake

AI tutoring could genuinely transform education. A tool that adapts to each kid's pace, available 24/7, patient and consistent — that's a dream teachers have described for decades. The Harvard study, limited as it is, shows it's possible.

But "possible" and "ready for national deployment" aren't the same thing. Edtech history is littered with tools that worked in controlled studies and flopped in real classrooms. One Laptop Per Child spent $2 billion distributing devices across developing countries. Learning outcomes barely moved. The devices gathered dust because nobody trained teachers to use them.

The pattern repeats: technology arrives, governments buy in, infrastructure lags, teachers feel bypassed, kids get a shiny tool that doesn't fix the underlying problem.

The Race Is On, Ready or Not

At least ten countries are now running national AI education programs. More are watching. The pressure to join is real — nobody wants to be the last country to "modernize."

But modernizing education isn't the same as deploying software. Estonia didn't become a top-performing education system by buying tech. It did it by investing in teachers, giving schools autonomy, and building curriculum over decades.

The countries racing to deploy AI tutors might get lucky. The technology might work despite the thin evidence. Or they might learn what every previous wave of edtech has taught: tools don't teach. Teachers do. And when you skip the evidence, students pay the price.