Gemini AI tutoring matches human performance in UK math study

Testing AI math tutoring in the UK

Google spent the past year working with researchers and educators in the UK to test how AI tutoring performs in real-world math lessons. The study involved over 1,100 students between the ages of 12 and 15 using a platform that combines Gemini's capabilities with specific pedagogical resources. The goal was to see if AI could identify a student's specific mistake and guide them to fix it independently.

To ensure the quality of the AI's responses, human experts reviewed every single interaction in real time before the student saw it. During the 10-week randomized controlled trial, students who received AI tutoring performed similarly to those who worked with human tutors on intermediate quizzes and daily tasks. This suggests that the technology can provide reliable, high-quality support when overseen by human educators.

The study also monitored whether students were using the AI to cheat or find shortcuts. The results showed that in nearly 99% of sessions, Gemini guided students through the logic instead of just handing over the final answer. Furthermore, students worked through the math problems themselves in more than 85% of their chats with the AI tutor.

Challenges in standardized testing

While the AI tutoring showed clear gains in daily classroom tasks, these improvements did not translate to higher scores on STAR standardized tests during the trial. Researchers believe this is because step-by-step tutoring takes more time per question than providing a simple hint. Consequently, students solved fewer total problems during their sessions compared to those using static hints.

Because the students spent more time on each individual problem, they covered less of the overall curriculum within the 10-week timeframe. The AI helped them understand the material more deeply and improve their pacing, but the trial was not long enough to cover the broad range of topics required for the standardized exams. This distinction highlights the difference between deep learning and rapid content coverage.

To address this, the research team is planning follow-up trials that will span a full academic year. They also intend to expand their metrics to better understand what specific factors drive long-term learning outcomes. These future studies will help determine how to balance the time-intensive nature of AI tutoring with the need to cover comprehensive curricula.

Transforming classrooms in Italy

In a separate study involving 31 schools in Italy, researchers looked at how Gemini for Education affected classroom dynamics and teacher wellbeing. They worked with over 1,600 teachers to collect data on lesson planning and student engagement. The findings indicated that AI is most effective when teachers remain at the center of the classroom, using the tool as a partner to co-design their materials.

When educators used Gemini to plan their lessons, they reported feeling more creative than usual, with 79% of participants noting this shift compared to only 30% without the tool. This collaborative planning led to a significant change in how classes were structured. Traditional front-of-the-room lectures dropped by 38%, making room for more workshops, group activities, and peer discussions.

Students responded positively to these active learning environments, with reported engagement rising from 77% to 84%. The shift toward collaborative learning nearly tripled the amount of student-led activity in the classroom. Students also reported debating diverse viewpoints more frequently and taking greater ownership of their own learning processes during these co-designed lessons.

Supporting teacher wellbeing

One of the most significant findings from the Italian study concerned the mental and physical load on educators. Teachers often face the heavy burden of relying entirely on their own energy to keep a room of students engaged. However, in lessons co-planned with Gemini, that reliance on personal energy dropped by 66%.

The study found that lessons where the activity design itself drove the engagement doubled, rising from 23.2% to 52.2%. By creating structured activities that carried their own momentum, teachers were able to reduce the daily pressure of constant engagement management. This allowed them to focus more on the human elements of teaching, such as mentoring and connecting with individual students.

Even on days with very little preparation time, the AI served as a thought partner to help teachers adapt their materials quickly. This support helped educators maintain high-quality instruction without the usual stress of last-minute planning. The research suggests that AI can act as a buffer against burnout by handling some of the heavy lifting involved in creative lesson design.