Google DeepMind's latest classroom trial in Sierra Leone shows that AI tutoring works best when teachers remain in control. A randomized study of 1,763 Grade 7 and Grade 8 students across 48 classrooms found that those using Google's Gemini Guided Learning during math lessons scored significantly higher than peers receiving standard instruction, with gains equivalent to roughly 1.2 to 1.7 years of typical learning progress in low- and middle-income countries.

What Exactly Did the Sierra Leone Study Test?

The trial ran from October to December 2025 across 12 government-supported junior secondary schools in Port Loko District. Teachers incorporated Gemini Guided Learning into two of their four weekly math periods, targeting about 12 hours of AI-assisted instruction over eight weeks. The study was preregistered, meaning researchers publicly committed to their methods before collecting data, a gold standard for research credibility.

Students worked in pairs on tablets or computers, with one student typing questions and the other taking notes and helping decide what to ask. Teachers remained responsible for setting lesson objectives, directing discussions, and providing extra help to struggling students. Local field monitors handled technical support but stayed out of classrooms during teaching.

The results showed an intent-to-treat effect of 0.258 standard deviations, which researchers estimated was roughly equivalent to moving a student from the 50th to the 60th percentile on a math assessment. Among students who completed at least the requested 12 hours of use, the effect was even larger at 0.380 standard deviations. Each additional hour of participation produced a 0.016 standard deviation improvement.

How Did Teachers Actually Use the AI Tool in Class?

All participating teachers completed a five- to six-hour training session covering tablet use, generative AI basics, the Gemini app, lesson preparation, and classroom management strategies. Teachers learned how to establish clear learning objectives, create starter prompts, and prepare question stems that students could use when stuck.

Lessons followed a consistent four-part structure designed to keep teachers central to the learning process:

• Introduction: Teachers introduced the lesson objective and learning goals to the whole class.
• AI-Assisted Work: Students worked in pairs with Gemini, with one student driving the conversation and the other navigating and taking notes.
• Class Discussion: The class came together to discuss what they had learned from their interactions with the AI.
• Teacher Closure: Teachers summarized the main points and reinforced key concepts.

Teachers also reported using Gemini to prepare lessons and find alternative explanations for familiar topics like fractions, ratios, and place value. Some said the technology reduced preparation time, though others found that reviewing different explanations from Gemini took longer than their previous approaches.

"This wasn't just about dropping the AI into the classrooms, it was a sociotechnical intervention. Teachers were supported to stay in the lead and use AI as a targeted tool that complements traditional instruction," said Irina Jurenka, Research Director at Google DeepMind.

Irina Jurenka, Research Director at Google DeepMind

What Did the AI Actually Do During Lessons?

Students and Gemini exchanged 113,344 messages across 7,421 conversations during the trial. Researchers analyzed these conversations and found that 97.4% of messages stayed on topic. About 91.4% of conversations focused on developing mathematical understanding or skills, while only 5% were primarily seeking a direct solution.

Gemini's approach was deliberately scaffolded, meaning it asked guiding questions rather than handing out answers. The AI posed scaffolding questions in 76.4% of its messages and supplied direct solutions in just 2.1%. This approach appeared to shift student behavior over time. In the first week, 67.7% of student interactions were skill-seeking, but that rose to over 90% in most later weeks, while solution-seeking fell from 25.1% to substantially lower levels.

Who Benefited Most From the AI Intervention?

One important finding emerged: students who entered the study with stronger baseline math skills recorded the largest gains. For every additional standard deviation in starting math proficiency, the treatment effect increased by an estimated 0.195 standard deviations. This suggests that while the intervention helped all students, it was particularly powerful for those already performing well.

Teachers participating in post-trial focus groups described strong student engagement. One teacher told researchers, "But the introduction of AI, I mean, let me confess, I've seen children rushing to attend classes." Teachers also identified barriers, including problems with traditional literacy, digital literacy, typing skills, and device access. Many asked for more time and wider access to help students use Gemini effectively.

What Happens Next for AI in African Classrooms?

The Sierra Leone study is the first in a planned international series of preregistered trials examining how Guided Learning affects teaching and learning in different education systems. Google DeepMind and Fab AI have released the teacher training materials and a rapid randomized controlled trial playbook covering the design, implementation, and analysis of the study, making the methodology available to other researchers and educators.

"We look to be innovative and improve service delivery, but we must also rigorously study the results of our innovations. I am therefore delighted that we now have strong evidence that carefully designed AI can help improve learning outcomes in support of our many hard-working teachers," said Conrad Sackey, Sierra Leone's Minister of Basic and Senior Secondary Education.

Conrad Sackey, Sierra Leone's Minister of Basic and Senior Secondary Education

The trial was authorized by the Sierra Leone Ministry of Basic and Senior Secondary Education and received approval from the Sierra Leone Ethics and Scientific Review Committee. Parental consent, teacher consent, and student assent were collected before participation. Oxford MeasurEd independently developed and administered curriculum-aligned assessments, while Laterite managed data collection and EducAid led local implementation.

Of the 871 students assigned to Guided Learning classrooms, 69% completed at least the requested 12 hours of use, with treatment classrooms averaging approximately 15 hours, 25% above the target level. The study's rigorous design, transparent methodology, and focus on teacher-led integration rather than AI replacement offer a roadmap for how AI tools might be responsibly scaled in resource-constrained education systems worldwide.