An AI maths tutor built by Google and the UK learning platform Eedi helped pupils as much as human tutors did during lessons in a 10-week randomised controlled trial in UK schools, but it made no measurable difference on a standardised maths test. Google, Eedi and Google DeepMind published the results on 29 October in a blog post, a release and a joint technical report.
The trial involved 1,179 pupils aged 12 to 15 in 10 UK state schools, according to Eedi. Every message the AI drafted was checked by a human tutor before a pupil saw it. Google presents the findings as showing AI tutoring can match human support; the technical report, written by Google’s LearnLM team and Eedi, is more cautious about what the null result on the broader test means.
In brief
- In a 10-week trial of 1,179 UK pupils, AI tutoring and human tutoring both lifted in-lesson accuracy to about 57%, against 43% with static hints.
- On the external Star Maths test, pupils given AI tutoring scored about the same as those given static content, and expert human tutoring did slightly worse.
- A separate Google study in Italy reports that teachers felt more creative using Gemini, but its figures are self-reported.
How the AI maths tutor trial worked
The AI maths tutor sits inside Eedi’s existing platform. Pupils there work through diagnostic questions designed to expose specific misconceptions. In this trial, each pupil was randomly assigned to one kind of help for the whole term whenever they got a question wrong: Eedi’s standard static content of hints and videos, an expert human tutor, or one of two versions of an AI tutor that used Gemini. One AI version saw only the current question and the likely misconception; the other also received a summary of the pupil’s past work.
The trial was pre-registered, meaning its design and analysis plan were published in advance, under the American Economic Association registry number AEARCTR-0018079. Eedi says it led the trial design, school recruitment and implementation, and built the constrained tutor, while the results were analysed and published jointly with Google DeepMind.
Supervising tutors reviewed 94,385 AI drafts and approved 94.7% without changes, the report says, up from 74.4% in the companies’ 2025 exploratory study. Google’s blog gives the earlier figure as 76%. The report says the AI held back the answer until the pupil had tried to work it out in 98.8% of conversations. Eedi says 7 AI messages containing a false maths statement reached pupils, 0.01% of the full corpus, against 7 genuine maths errors in human tutors’ messages, or 0.02%.
Strong in the lesson, flat on Star Maths
On the immediate measure, tutoring of either kind helped. After a mistake and some support, pupils answered a similar follow-up question correctly 57.7% of the time with a human tutor and 56.9% and 56.5% with the two AI versions, against 43.4% with static content, according to the report.
The broader test told a different story. Star Maths, an external standardised assessment from Renaissance, was taken before and after the trial; Eedi says 916 pupils sat both. Pupils given the question-only AI tutor scored 0.01 standard deviations lower than those on static content, and those given the personalised version 0.02 standard deviations higher: in effect, no difference. Expert human tutoring came out 0.09 standard deviations lower, though Eedi says that gap could be due to chance.
Google’s explanation is time. Step-by-step tutoring takes longer than reading a hint, so pupils solved about 14 problems per session with AI tutoring against 19 with static hints, the blog says. The report puts the gap over the term at a median of 99 study units for AI-tutored pupils and 77 for those with human tutors, against 132 for static content. Fewer units meant fewer topics covered on a test that spans the whole curriculum.
An exploratory analysis in the report found a gender pattern: on static content, boys pulled ahead of girls on Star Maths by 0.31 standard deviations, and all three forms of tutoring reduced that gap to near zero. Eedi calls these results tentative.
The Italy figures are self-reported
Google’s post also describes a year-long study with the Salesians of Don Bosco school network and the Pontifical Salesian University in Rome, across 31 schools and 1,600 teachers representing more than 24,500 students aged 13 to 18. It was based on classroom activity records and self-reports. In those self-reports, 79% of teachers said they felt more creative than usual when planning with Gemini, against 30% without it. Google says front-of-class lecturing fell by 38%, and student-reported engagement rose from 77% to 84%. The study compared lessons planned with and without Gemini rather than randomly assigning pupils, as the UK trial did.
Why ten weeks of better lessons did not move Star Maths
The honest reading is that the trial answered one question well and left the more important one open. It shows, with a pre-registered design, that a supervised AI maths tutor can match trained human tutors at the moment a pupil gets something wrong, and that it can do so with very few errors reaching children. That is a real result, and rarer than it should be: most claims about AI in classrooms rest on pilots without comparison groups.
What it does not show is that pupils learned more maths overall. On the measure that schools and parents care about, neither AI nor human tutoring beat static hints over a term. The coverage explanation is plausible, and the report backs it with platform data, but it is the developers’ own account of their own null result. Eedi says its earlier two-year trial of the platform looked similar at 10 weeks and only showed clear gains after a year or more, which is why it has launched a year-long trial with more than 15,000 pupils in 26 state schools.
The cost of running the system is also unclear. Every AI message in this trial was approved by a human tutor; Eedi says its next trial moves to tutors overseeing conversations rather than approving each message. Until that larger study reports, the evidence supports an AI maths tutor as a safe, supervised substitute for a human tutor in the moment, not as a proven way to raise attainment.
That pattern of a strong early signal and a missed main goal is familiar from medicine, as with this week’s KRAS pancreatic cancer vaccine trial. It also fits a wider point TSN has made about AI tools: they are already in the toolbox but still fall short of finishing the job on their own. Google’s education push also includes free AI Quests for pupils aged 11 to 14, which have no outcome data yet. More evidence-based stories are collected in our weekly tech for good round-up.
Sources
Primary:
- Google blog, Kevin McKee and Amanda Rosenburg, “How Gemini is helping students and teachers in the classroom”, 29 October 2026: https://blog.google/products-and-platforms/products/education/impact-studies-uk-italy/
- Eedi, “Eedi Publishes Full Findings from Its Most Recent AI Tutor Trial”, London, 29 October 2026: https://www.eedi.com/news/eedi-publishes-full-findings-from-its-most-recent-ai-tutor-trial
- LearnLM Team, Google & Eedi, “Measuring the impact of AI tutoring, from near transfer to standardised tests: An expanded RCT in UK mathematics classrooms”, technical report, 29 October 2026: https://services.google.com/fh/files/misc/learnlm_oct26.pdf
Press: none used.
Related stories
- Google’s Third Free AI Quest Has Pupils Aged 11 to 14 Map a Brain
- Already in the Toolbox: AI Is Here, and It Still Doesn’t Finish the Job
- DeepMind’s Genie 4 Adds Sound to Worlds That Last a Few Minutes
- KRAS Pancreatic Cancer Vaccine Misses Main Goal but Primes T Cells
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
