A human-supervised AI maths tutor matched expert tutors in a 1,179-student UK randomised trial: 56.9% right on the follow-up question versus 57.7%

In a pre-registered randomised trial of 1,179 UK secondary students, an AI maths tutor checked by human educators matched expert human tutors at helping students fix mistakes: 56.9% answered the follow-up question correctly versus 57.7%, and 43.4% with standard hints. After 10 weeks, no tutoring type beat standard help on a broad standardised test.

When a student gets a maths question wrong on Eedi, a platform several UK schools use in their maths lessons, the usual help is a short hint or a video. A live tutor who chats with the student works better, but human tutors are expensive, and Eedi's Chief Impact Officer says most schools can't afford them. So Eedi and Google DeepMind asked whether Gemini could do the chatting, with a person watching.

The clever bit

When a student missed a question, Gemini drafted a chat reply that knew the question, and in one version the student's learning history too. It was told to guide the student step by step rather than hand over the answer. One of 58 expert maths educators then approved, regenerated or rewrote each draft before the student saw it. They approved 94.7% of 94,385 reviewed drafts unchanged, up from 74.4% in a 2025 pilot.

The trial was registered in advance and ran across 10 UK state schools. Each of 1,179 students in Years 8 to 10 was randomly assigned to standard hints, an expert human tutor, or one of two AI tutors for 10 weeks. After getting help, 43.4% of students with standard hints answered the follow-up question correctly, against 57.7% with an expert tutor and 56.9% and 56.5% with the two AI tutors. On the first question of the next topic, standard help scored 50.5% and both AI tutors scored 58.4% and 58.3%.

What we saw was that in the lesson, the AI Tutor performed as well as a human tutor, and did so safely.

— Bibi Groot, Chief Impact Officer, Eedi, via eedi.com

Hooray

The safety audit backs her up. Maths errors showed up in fewer than 3 of every 1,000 AI messages, close to the expert tutors' rate, and Eedi reports that no harmful content reached a student. On a quiz partway through the trial, AI tutoring also edged ahead of standard help, 63.5% against 59.6% for the personalised version.

What comes next

On Star Maths, an independent standardised test many UK schools use, nothing moved: after 10 weeks AI-tutored students scored about the same as students who got standard help. The report points to a trade-off, since a chat takes longer than a hint and tutored students covered fewer topics. A year-long follow-up that began in September includes more than 15,000 students in 26 UK schools, with humans stepping in only when a conversation goes off track.

How solid is it?

Promising: Results measured in a study or bounded trial. Limitations and wider testing still matter.

Pre-registered randomised controlled trial (AEA RCT Registry AEARCTR-0018079) run by Eedi and Google DeepMind's LearnLM team in 10 UK state secondary schools: 1,179 students in Years 8 to 10, each randomly assigned to standard remediation, an expert human tutor, or one of two Gemini-based AI tutors with a human educator approving every message, over 10 weeks of weekly lessons. Follow-up question correct: 43.4% standard, 57.7% expert tutor, 56.9% and 56.5% AI. Next topic's first question: 50.5% standard, 59.1% expert, 58.4% and 58.3% AI. The 74 (Greg Toppo, Oct 29, 2026) reported the results independently and interviewed Eedi's Chief Impact Officer.

See for yourself.

Selected for Nov 3, 2026. Reviewed Nov 3, 2026, 5:58 AM MT. Source dates may be earlier than this edition.