The UK Is Spending £4 Per Student on AI Tutors. A Human Tutor Costs £38 an Hour.

England just committed £1.8 million to build AI tutoring tools for 450,000 disadvantaged students. Divide those numbers and you get roughly £4 per kid.

A private tutor in England costs £30 to £50 an hour.

That's the bet Education Secretary Bridget Phillipson is making: that artificial intelligence can deliver what money never could — a personal tutor for every struggling child in the country.

The numbers that forced the decision

The 450,000 figure isn't random. It's the number of students in Years 9 through 11 who receive free school meals — the UK government's proxy for economic disadvantage. These are kids heading toward their GCSEs, the exams that determine whether doors open or close.

England tried the human version of this already. The National Tutoring Programme, launched after COVID, spent hundreds of millions connecting disadvantaged students with real tutors. By its final year in 2023-24, the programme had reached millions of sessions. But it was expensive, hard to scale, and quality varied wildly depending on the tutor.

Now the Department for Education wants to try something different. It's seeking a private sector partner for a contract running nearly three years, with tools expected in schools by late 2027. Teachers will co-design the AI tools starting this summer. If it works, it goes national.

"AI tutoring tools could level the playing field for families who cannot afford private assistance," Phillipson said. The emphasis on could is doing a lot of work in that sentence.

What the research actually says

Here's the thing — the evidence is surprisingly strong.

A randomized controlled trial published in Scientific Reports found that students using AI tutors learned more in less time than students in traditional active-learning classrooms. They also reported feeling more engaged and more motivated.

A meta-review by Kulik and Fletcher covering 50 studies concluded that intelligent tutoring systems can "match the success" of human tutoring. Not outperform. Match.

And a Brookings Institution report from January 2026 looked at four recent randomized controlled trials and found "substantial learning gains across all studies," along with better knowledge transfer, improved motivation, and greater efficiency.

The Tutor CoPilot trial — one of the largest — found that students whose tutors used an AI assistant achieved higher mastery rates, with the biggest gains among the least experienced tutors. The AI didn't replace the teacher. It made mediocre tutoring good.

That last finding matters most for England's plan. The government isn't trying to remove teachers from the equation. It's trying to give every classroom an invisible teaching assistant that never sleeps, never gets tired, and adapts to each student's pace in real time.

The £4 question

Let's go back to that £4 per student figure. It sounds absurd. But it's also misleading.

The £1.8 million is for development — building the tools, testing them, iterating with teachers. Once built, the marginal cost of serving each additional student drops toward zero. That's the entire promise of software. You build it once, deploy it millions of times.

Compare that to human tutoring. At £38 per hour (the UK average), giving each of those 450,000 students just 10 hours of personal tutoring would cost £171 million. Twenty hours? £342 million. And that's assuming you could find enough qualified tutors, which you can't.

The UK tried. The National Tutoring Programme struggled constantly with recruitment and quality. Rural areas got fewer tutors. Low-income areas got less experienced ones. The students who needed the most help often got the least qualified support.

AI doesn't have a supply problem. It scales instantly. The question isn't whether it's cheaper — it's whether it's good enough.

What could go wrong

Plenty.

Critics point to real risks. AI tutors can generate wrong answers with total confidence. They can miss when a student is confused in a way that a human would catch from body language. They can't handle the emotional side of learning — the kid who's struggling not because they don't understand fractions but because their parents are splitting up.

There's also the alignment problem. The Department for Education says all tools will be curriculum-aligned and safety-tested. But building AI that teaches the UK maths curriculum accurately, adapts to different learning styles, catches mistakes in real time, and doesn't occasionally hallucinate — that's a hard engineering problem. £1.8 million is a modest budget for solving it.

And then there's the dependency question. A study published in PNAS by Bastani et al. found that students who relied heavily on AI tutoring showed learning gains while using the tool but struggled more when it was removed. The AI taught them to lean on the AI, not to think independently.

The global race

England isn't alone in this bet.

Massachusetts announced a partnership with Google last week to offer free AI training to all state residents — the first state in America to do so. Governor Maura Healey wants to make it the "applied AI capital" of the US.

Japan launched a student-led nonprofit offering low-cost online tutoring to reduce education inequality among elementary and middle school students. Tokyo schools reported a surge in uptake.

Khan Academy's Khanmigo AI tutor jumped from 40,000 to 700,000 K-12 students in a single academic year, with projections to pass one million in 2025-26. Sal Khan has been arguing for years that AI could finally deliver the "two sigma" effect — Benjamin Bloom's 1984 finding that one-on-one tutoring produces learning gains two standard deviations above classroom instruction.

Bloom's problem was that one-on-one tutoring for every student was impossible. It was too expensive, too labour-intensive, too dependent on finding enough good tutors.

AI might be the answer to Bloom's problem. Or it might be a very expensive way to discover that teaching is harder than pattern recognition.

What happens next

The UK's pilot runs through the next academic year. Performance data will track student progress, engagement, and teacher feedback. If the results hold, nationwide rollout could begin by 2027.

The contract closes in May. Teachers start co-designing tools this summer. The first students could be using AI tutors by autumn.

Four hundred and fifty thousand kids are waiting. Most of them have never had a personal tutor. Most never will — unless this works.

The price of finding out: £4 per student. The cost of not trying: another generation where your postcode determines your grades.