Google DeepMind Deploys Gemini Guided Learning in Sierra Leone: +0.258 SD Math Score Gain

Title

Google DeepMind Deploys Gemini Guided Learning in Sierra Leone: +0.258 SD Math Score Gain

Content

Google DeepMind ran a pre-registered randomized controlled trial in Sierra Leone schools deploying Gemini Guided Learning, in partnership with EducAid, Laterite, Oxford MeasurEd, and with support from Google.org and the Gates Foundation. Across the eight-week trial, students using Guided Learning saw a gain of +0.258 standard deviations in math scores compared to the control group, representing roughly 1.2 to 1.7 years of typical learning progress in just two months. Classrooms where teachers used Gemini for about half of their lessons to meet a 12-hour target saw even higher gains, roughly 1.8 to 2.5 years of progress. "The results from this pre-registered trial suggest that AI can be a powerful pedagogical partner — not by replacing teachers, but by augmenting their reach," DeepMind stated. Engagement was remarkably high: 69% of students met or exceeded usage targets, far surpassing the 5% typical for voluntary educational technology, often called "The Five Percent Problem." Teachers reported that Gemini supported their own professional growth, with many shifting from "lecturers" to "facilitators." The motivation was concern that generative AI could become a shortcut for students, bypassing the cognitive effort required for deeper learning. Guided Learning was built from years of research and LearnLM work to be pedagogically grounded and tuned to prioritize building understanding over providing direct answers. An analysis of over 113,000 interactions exchanged during the trial revealed that students used the tool to build conceptual understanding in 91.4% of conversations, rather than seeking solutions. Gemini responded by posing scaffolding questions in 76% of its messages, providing direct solutions in only 2% of cases. Technically, the system uses Gemini fine-tuned via LearnLM for education, designed to deliver Socratic-style scaffolding. Students' conversations became more learning-oriented over time: skill-building queries rose to 90% by the final week, up from 68% in the first week, while solution-seeking questions dropped from 25% to 10%. The trial ran for eight weeks in classrooms with students who used the tool for about half of lessons, with the treatment vs. control design enabling the SD measurement. DeepMind worked with Fab AI on protocols and is releasing a teacher training guide and an RCT playbook. Looking forward, the trial also highlighted an achievement gap: students entering with stronger math skills benefited most. DeepMind plans to expand the trials to other countries and probe more deeply into metacognition and relational intelligence, with support of the Global AI for Learning Alliance (GAILA) to accelerate commitments and build a cross-country evidence base.