Google DeepMind, in partnership with Fab AI and the Sierra Leone Ministry of Education, has published results from a randomized controlled trial (RCT) evaluating the impact of Gemini’s Guided Learning feature on junior secondary school math outcomes. The eight-week study involved 1,763 students across 12 schools in Port Loko District and represents one of the more rigorously structured evaluations of generative AI in a classroom setting to date.
Quantitative Outcomes
Students in the Guided Learning group recorded a gain of +0.258 standard deviations in math scores relative to the control group, which the researchers translate to approximately 1.2 to 1.7 years of typical learning progress within the trial period. In classrooms where teachers integrated Gemini into roughly half their lessons, targeting 12 hours of use during the trial, gains rose further, representing an estimated 1.8 to 2.5 years of progress. Student engagement was notably high: 69% of participants met or exceeded usage targets, a figure that contrasts sharply with the roughly 5% voluntary adoption rate commonly observed for educational technology tools.
Pedagogical Design and Interaction Patterns
A central concern with generative AI in education is that students will use it to bypass genuine cognitive effort. DeepMind’s Guided Learning is explicitly designed to counter this tendency. Analysis of more than 113,000 student-AI interactions showed that students used the tool to build conceptual understanding in 91.4% of conversations. Gemini posed scaffolding questions in 76% of its responses and provided direct answers in only 2% of cases.
Behavioral trends over the trial period reinforced this pattern. Skill-building queries rose from 68% in the first week to 90% by the final week, while solution-seeking queries dropped from 25% to 10%. The interaction model is described as Socratic, keeping the cognitive work with the student rather than offloading it to the model.
Teacher Role and Professional Development
Teachers were not sidelined in the trial design. Educators set lesson objectives, designed activities, and led classroom discussions. In focus groups, teachers reported that using Gemini for lesson preparation helped them discover new ways to explain topics such as fractions, with many describing a shift in their own practice from direct instruction toward facilitation.
DeepMind is releasing a teacher training guide developed with Fab AI, including the protocols used during the study, to support replication in other contexts.
Limitations and Next Steps
The study also surfaced a notable caveat: students who entered the trial with stronger existing math skills benefited the most, indicating a persistent achievement gap. DeepMind acknowledged that tools need to be refined to deliver stronger gains for lower-performing students.
Additional pre-registered RCTs are planned across multiple countries. DeepMind is also releasing a methodology playbook for running scalable studies, and cited its support of the Global AI for Learning Alliance (GAILA) as part of a broader effort to build a cross-country evidence base for AI in education.