Google’s AI math tutor matched human tutors in a UK trial, but test scores stayed flat

Google has published results from two classroom studies of Gemini. In a 10-week UK trial with over 1,100 students, Google says an AI math tutor performed similarly to human tutors on in-platform tasks and quizzes. Standardized test scores did not improve.

A 10-week trial with human review of every reply

The UK study, run with the education platform Eedi, was a randomized controlled trial with students aged 12 to 15. When a student answered a question wrongly, they received one of three kinds of help: a static hint, a human tutor, or interactive AI tutoring that combined Eedi’s teaching resources with Gemini. Human experts reviewed every AI response in real time before it reached the student, so the result describes supervised tutoring, not a chatbot left alone with children.

Google reports that expert reviewers fully approved 95% of Gemini’s responses, up from 76% in last year’s study, which it describes as a pedagogical quality gain of nearly 25%. Following a prompt Google supplied, Gemini guided students instead of giving answers in nearly 99% of sessions, and students did the math themselves in more than 85% of chats.

The gains did not carry over to the independent STAR standardized tests. Google’s explanation is that step-by-step tutoring is slower: students solved about 14 problems per session with AI tutoring against 19 with static hints, so they covered less of the curriculum in 10 weeks. Google says it is planning follow-up trials across a full academic year.

Italy: More group work, less pressure on teachers

The second study was a year-long project run together with the Don Bosco Salesian school network and a Salesian pontifical university based in Rome. It covered 31 schools and 1,600 teachers, representing over 24,500 students aged 13 to 18. Teachers reported on lessons planned with and without Gemini for Education.

According to Google, 79% of teachers felt more creative when co-designing with Gemini, against 30% without it. Front-of-class lecturing fell by 38%, and student-led collaborative learning nearly tripled. Student-reported engagement rose from 77% to 84%. Lessons where the activity design itself carried engagement rose from 23.2% to 52.2%, while teachers’ reliance on their own energy fell by 66%. These figures come from self reports and collected classroom activities.

Numbers in context: A bigger, tougher test than last year’s

The new trial is much larger than the first one. The earlier exploratory study, according to Eedi’s write-up and a preprint on arXiv, ran in summer 2025 in five UK secondary classrooms with 165 students over seven weeks. On the next related question, that study found 66% success with the AI tutor, 60% with a human tutor and 56% with a static hint. The new trial has roughly seven times as many students and adds an external measure, the STAR tests, which the earlier work did not report. That is where the headline gap appears.

The trade-off Google describes matters for schools. Less volume per session may be a fair price for deeper work, but a 10-week window cannot show it. Until the full-year trials report, the claim that AI tutoring improves long-term outcomes remains unproven, and Google itself does not make it.

The student side of Google’s education push is also moving. Our earlier report on Gemini’s zoomable explorations in the student hub noted that Google gave no learning-outcome data for that feature, which makes these controlled studies the first outcome evidence it has offered in this series.

What the studies leave unanswered

The announcement links technical and implementation reports, but the page does not say whether the work has been peer reviewed. All figures are Google’s and its partners’ own. The tutoring ran with a custom prompt, Eedi’s platform and live expert review, so it cannot be assumed that classroom use of Gemini would behave the same way unsupervised. The Italian findings rest largely on teacher and student self reports rather than test results. The page also does not say when, or whether, the AI tutor will be offered to schools beyond the trial, or at what price.