Google finds classroom gains from Gemini tutoring, but no standardized-test gains

Students receiving Gemini-powered math tutoring performed similarly to those receiving human help on classroom tasks and quizzes—but saw no gains on independent standardized tests. Google published the findings on October 29, 2026, alongside Italian classroom research. A crucial condition shaped the UK results: human experts checked every AI response before students saw it.

A tutor with a human checkpoint

The UK study was a 10-week randomized controlled trial involving more than 1,100 students aged 12–15. It used Eedi’s school platform to compare three kinds of help after a wrong answer: a static hint, a human tutor, or interactive AI tutoring combining Gemini with Eedi’s teaching resources.

Teachers monitored progress and supported students in the classroom, while experts reviewed AI replies in real time. The results therefore describe tutoring delivered with that oversight, rather than an unchecked chatbot working alone. Google says reviewers approved 95% of responses outright, up from 76% in its previous study.

The AI was instructed to guide students through problems instead of supplying answers. Google says it followed that approach in nearly 99% of sessions, and students worked through the math themselves in more than 85% of chats. Those measures describe how the tutoring worked, separately from whether learning gains carried over to broader tests.

More guidance, fewer problems completed

Researchers tracked regular platform assignments, intermediate quizzes and independent STAR standardized tests. Google reported gains on classroom tasks and quizzes, with AI-supported students performing similarly to the human-tutored group. But those gains did not appear in STAR scores during the trial.

Google’s proposed explanation is a tradeoff between guidance and coverage. Step-by-step tutoring took longer than simple hints, so students completed fewer problems and covered less curriculum material. The company believes 10 weeks was too short to cover the range of topics assessed by STAR; that remains its explanation, not a demonstrated longer-term benefit.

Problems completed per session

01About 14

AI tutoring

Students receiving AI tutoring averaged about 14 problems per session.

02About 19

Static hints

Students receiving static hints averaged about 19 problems per session.

To examine that possibility, Google says it is exploring follow-up trials lasting a full academic year. It also plans to expand its measures of learning and outcomes, rather than rely only on the shorter trial’s results.

In Italy, teachers used Gemini before the lesson

The Italian study examined a different use: teachers co-designing lessons with Gemini for Education. Over a year, Google worked with the Salesians of Don Bosco school network and Pontifical Salesian University across 31 schools. The study involved 1,600 teachers representing more than 24,500 students aged 13–18.

Researchers collected classroom activities and self-reports about lessons planned with and without Gemini. Unlike the UK’s randomized tutoring trial, this work focused on classroom dynamics, engagement and teacher wellbeing. Google reported several shifts in lessons co-designed with the tool:

• Teachers reported feeling more creative in planning and teaching: 79% with Gemini, compared with 30% without it.
• Traditional front-of-class lectures fell by 38%, while student-led collaborative learning nearly tripled through workshops, group activities and peer discussions.
• Student-reported engagement rose from 77% to 84%, alongside more active participation in lessons.

Google also reported a change in what kept classes engaged. Teachers’ reliance on their own energy to sustain engagement dropped by 66%. Lessons in which the activity design itself drove engagement rose from 23.2% to 52.2%, which the company described as easing teachers’ daily pressure. These are findings about lesson design and reported classroom experience, distinct from the UK study’s measured math outcomes.