AI Education Gap: 273 Million Children Still Out of School

Three reports landed the same week. Together they tell a story nobody planned: AI's making education better for children who already have it, and irrelevant for the 273 million who don't.

UNESCO's 2026 Global Education Monitoring Report found out-of-school numbers climbed for a seventh straight year. The same day, Anthropic's fifth economic impact index showed AI tools concentrate benefits in high-income countries among experienced users. A Stanford review of 800+ studies on AI in classrooms concluded only 20 met rigorous causal standards — and learning gains vanish when the AI's taken away.

Same finding, three angles.

The number keeps climbing

One in six school-age children isn't in school. The 273 million figure — children, adolescents, and youth — has risen 3% since 2015, reversing decades of progress. Sub-Saharan Africa's hit hardest: population growth outpacing school construction. Conflict zones push the real number higher. UNESCO estimates 13 million more children in the ten worst-affected countries aren't captured in official stats.

The Middle East's become a crisis point. The Iran war has disrupted education for 52 million children. Schools across multiple countries have shut. Millions may never return.

What makes the 2026 data different: it arrives the exact moment governments are betting billions on AI as the fix.

The $55,000 AI classroom

Alpha Schools opens its newest campus in Chicago this fall. Children learn core academics from AI software for two hours each morning. No teachers — only "guides" who motivate and coach. The company's grown from 200 students to over 1,000 at 22 locations, with plans for 35.

Annual tuition: $55,000.

The most advanced AI education available — the version that replaces teachers entirely — costs roughly what Stanford does. The children UNESCO's counting, in sub-Saharan Africa and conflict zones and rural South Asia, aren't just priced out. Many don't have electricity.

UNESCO's data makes this concrete. Electrification alone's been linked to nearly a full extra year of schooling in Cambodia. School feeding programs add up to half a year of learning for every $100 spent in low- and middle-income countries. Basics still matter more than algorithms.

What Stanford actually found

Stanford's AI Hub for Education reviewed 800+ academic papers on AI in K-12 classrooms. Only about 20 were high-quality causal studies — the kind that prove whether AI itself drives outcomes, not teacher quality or student motivation.

Results were specific and limited. AI helped students on structured tasks — maths practice, writing feedback. Teachers saved up to 30% of grading and planning time without lowering quality.

But when the AI was removed, gains disappeared.

"It's still too early to know if this is doing school faster or if it's reimagining," said Chris Agnew, managing director of Stanford's AI Hub. "The headline is mixed."

Lily Fesler, a senior researcher, added that without causal evidence, improvements could be driven by teacher experience or classroom context rather than the tool itself. Most existing research reflects simple chatbot-style products, not real classroom use.

At least ten countries — El Salvador, Kazakhstan, Estonia, the UAE — are already deploying AI tutoring at national scale. Running ahead of the evidence.

The Anthropic paradox

The most striking data came from the company that should be most optimistic about AI's potential.

Anthropic's fifth economic impact report analysed how Claude's used across the economy. The education findings are quietly damning for the "AI as equaliser" narrative.

First: AI tools are used far more intensely in high-income countries. The top 20 account for 48% of per-capita usage, up from 45%. Global inequality in AI adoption's growing, not shrinking.

Second: experienced users get dramatically more value. Six-month-plus users have a 10% higher success rate — a gap not explained by task choice or location. They've learned to use AI effectively. That learning compounds.

Third: as adoption spreads, average economic value per session is falling. New users try lower-value tasks. Power users pull further ahead.

Anthropic's head of economics, Peter McCrory, said the pattern means "displacement effects could materialise very quickly." His recommendation: build monitoring systems now, before the gap becomes permanent.

The implication for education's clear. AI tutoring works better the more you already know how to use AI. Children in well-resourced schools with trained teachers and reliable internet will improve. Children without those foundations won't just miss out — they'll fall further behind than they would have without AI.

The evidence we're missing

Nobody's published results from a national-scale AI tutoring deployment in a developing country.

El Salvador handed its entire public system to xAI's Grok — one million students across 5,000 schools. OpenAI's "Education for Countries" runs in eight nations. Google plans to train six million teachers on AI. Real experiments, running now.

None have released outcome data. Not one. The LLMs-in-education market's projected to grow from $7.5 billion to over $35 billion by 2030. Billions flowing in, zero causal evidence flowing out.

The AFT launched its National Academy for AI Instruction on March 18, training 400,000 teachers on AI classroom use — one of the few programmes acknowledging the technology doesn't work without the teacher. A Jobs for the Future survey found 40% of learners cite premium AI tool costs as a major barrier. In UNESCO's 273 million countries, that number approaches 100%.

What would actually help

UNESCO buries its most useful finding in the policy section. In 14 African countries, making education compulsory (not just free) added over a year of schooling. Child labour laws grew the gains further. Cash transfers tied to attendance made children 36% more likely to enrol.

Boring interventions. Nobody keynotes at Davos about school feeding. But they have causal evidence — the standard 98% of AI education studies fail to meet.

AI can probably help education. In specific contexts, for specific tasks, with trained teachers and working infrastructure. The question's whether the rush to deploy AI is pulling attention and funding from interventions that work now for the children who need them most.

273 million children aren't waiting for a better chatbot. They're waiting for a school.