Google published findings from two classroom studies on October 29, 2026. In a 10-week UK trial, students receiving supervised Gemini tutoring scored close to students receiving expert tutoring on near-term measures, while the overall Star Maths analysis found no advantage over standard Eedi remediation. A separate yearlong Italian study reported higher engagement in lessons co-designed with Gemini.
The UK trial compared four kinds of math support
The randomized trial assigned 1,179 students aged 12 to 15 across 10 state-funded secondary schools and 58 math classes. Students received standard Eedi remediation, expert tutoring, or one of two Gemini tutoring approaches. The standard-remediation group was an active comparison: students used videos, fluency exercises, and educator-written hints, rather than receiving no extra support.
Gemini drafted tutoring messages, but expert educators reviewed each draft before it reached a student. Tutors could approve, regenerate, or edit a message. Of 94,385 reviewed drafts, 94.7% were approved unchanged.
The two AI groups differed in the context available to Gemini. One used the question a student had missed; the other also used information about the student’s learning history. A platform bug kept summaries of earlier tutoring chats out of that history, which still included Eedi performance, pace, proficiency, and misconceptions. The trial also used three Gemini versions: Gemini 3.1 Pro until April 21, Gemini 3 Flash from April 21 to May 21, and Gemini 3.5 Flash from May 21 through the trial’s end.
Near-term scores were close; Star Maths showed no overall advantage
The technical report from Google and Eedi separates immediate correction, transfer to the next Eedi unit, an eight-question midtrial quiz, and Renaissance Star Maths, a standardized assessment. The first three measures track performance on platform work and recently taught material; Star Maths measures a broader outcome.
| Study condition | Immediate correction (% correct) | Next-unit transfer (% correct) | Midtrial quiz (% correct) | Star Maths effect vs. standard remediation (standard deviations) |
| Standard remediation | 43.4% | 50.5% | 59.6% | Reference |
| Expert tutoring | 57.7% | 59.1% | 61.8% | −0.09 (95% credible interval: −0.25 to +0.07) |
| Gemini with question context | 56.9% | 58.4% | 62.1% | −0.01 (95% credible interval: −0.15 to +0.14) |
| Gemini with personalized context | 56.5% | 58.3% | 63.5% | +0.02 (95% credible interval: −0.12 to +0.16) |
On immediate correction, next-unit transfer, and the quiz, the Gemini groups’ scores were close to the expert-tutoring group’s. The Star Maths estimates tell a different part of the story: all three 95% credible intervals include zero, so the full-sample analysis did not establish an overall advantage over standard remediation for either Gemini condition or expert tutoring.
Students completed a median of 99 Eedi study units in each Gemini group and 77 with expert tutoring, compared with 132 under standard remediation. Google’s report offers reduced topic coverage as one possible explanation for the Star Maths result and also describes exploratory gender-related patterns.
Italy’s findings focus on engagement and teaching practices
Google reported that its yearlong Italian study covered 31 schools, 1,600 teachers, and more than 24,500 students aged 13 to 18. The study drew on classroom activities and self-reports about lesson planning with and without Gemini for Education.
Student-reported engagement was 84% in lessons co-designed with Gemini, compared with 77% in the comparison lessons. Google also reported that 79% of teachers in the Gemini condition described greater creativity in teaching and planning, compared with 30% in the comparison; traditional front-of-class lectures were 38% lower, and student-led collaborative learning was nearly three times as common. These findings describe reported engagement and teaching practices rather than standardized academic-attainment results.
