Gemini as a Tutor: Google DeepMind Study Shows a Year's Math Gains in Eight Weeks

Google DeepMind has published the results of a randomized controlled trial in which hundreds of students in Sierra Leone used Gemini's Guided Learning mode as a math tutor. The effect turned out to be larger than the company expected, though the researchers themselves caution that the results should be read carefully.

Guided Learning isn't just a chatbot handing out ready-made answers. Google redesigned Gemini to respond to students with guiding questions rather than immediate solutions. In practice, the model replied with a follow-up question in 76 percent of messages, while giving a direct solution in just 2 percent of cases.

How Guided Learning works

An analysis of the conversations showed that students increasingly sought to understand the material rather than just get an answer. The share of skill-building questions rose from 68 percent in the first week to 90 percent in the last, while "just give me the answer" style questions fell from 25 to 10 percent.

Teachers at the participating schools received just one day of training on the tool, yet some said they themselves picked up new ways of explaining familiar topics, such as fractions, from the AI. Many described their role shifting from lecturer to learning companion, spending more time on one-on-one conversations with students.

Sierra Leone's government reacts

We have to be innovative and improve the quality of our services, but we also have to rigorously study the effects of our innovations. So I'm glad that we now have strong evidence that well-designed AI can improve learning outcomes by supporting our hardworking teachers - Conrad Sackey, Sierra Leone's Minister of Basic and Senior Secondary Education

For Sierra Leone, a country with a limited pool of qualified math teachers, the result matters practically as well as academically. The Ministry of Education took part in the project as a co-author of the study, not merely its subject, which points to growing trust among local authorities in AI tools for education.

Caveats from the researchers

Google DeepMind acknowledges the study has limitations. Students who were already stronger in math gained the most, while weaker students showed smaller improvements. The study's authors say plainly that the "additional year of learning" estimate should be taken with a grain of salt, since converting a standard deviation into years of education is a simplification.

The team has already announced talks with organizations that specialize in improving outcomes for lower-performing students, since the current pedagogical model may not be optimal for them. The methodology, teacher guide and research playbook have been made public so other teams can replicate the experiment in other countries.

Relevance for Polish schools

In Poland, the debate about AI in education focuses mainly on detecting chatbot-written homework, less often on using AI as a tutor that supports learning. The Sierra Leone result points to a different direction: a model designed to guide rather than hand out answers can produce a measurable effect even with minimal teacher training and a short rollout period.

For Polish schools, especially smaller and rural ones lacking math teachers, the model tested in Sierra Leone could serve as a reference point, though the educational contexts of the two countries differ substantially in resources and internet access. Google's decision to make the methodology public, however, makes it easier for Polish local governments or education organizations to test similar tools.

Sources: Measuring the impact of learning with AI in Sierra Leone and beyond (deepmind.google), Google Tested Its AI Tutor In Real Classrooms. It Worked (forbes.com), Measuring the impact of AI on teaching and learning (blog.google)