Inside Google DeepMind’s First Randomized Trial on Guided AI Learning in Sierra Leone

The promise of AI in education is immense, but rigorous evidence of its impact in real classrooms has been scarce. In June 2026, Google DeepMind and the educational non-profit Fab AI published results from a landmark randomized controlled trial (RCT) in Sierra Leone, offering one of the most detailed looks yet at how a structured AI learning tool can work alongside teachers .

The study found that integrating Gemini's "Guided Learning" mode into math lessons significantly improved student test scores. However, the results also illuminated critical nuances: students who were already ahead in math benefited the most, teachers had to adapt their roles, and the carefully designed intervention is difficult to separate from its supporting components.

How the Trial Was Designed

Conducted in partnership with the Sierra Leone Ministry of Education, the trial was a preregistered, two-arm cluster-randomized controlled trial. It ran for eight weeks, from 6 October to 5 December 2025, in 12 government-supported junior secondary schools in the Port Loko District .

The study enrolled 1,763 students in Grades 7 and 8 (aged 13 and older) across 48 mathematics classrooms . Within each school and grade, classrooms were randomly assigned to either the treatment group, which used Gemini Guided Learning, or the control group, which continued with standard instruction. To avoid confounding the effect of the AI tool with teacher quality, all teachers received the same 5–6 hours of training .

The target dosage was for students to use the AI tool during two of their four weekly math periods, equalling 12 hours of Guided Learning over the eight weeks .

A Structured Four-Part Lesson with a Twist

The intervention wasn't simply handing students a chatbot. It followed a strict pedagogical framework designed to make the AI a partner in a collaborative learning process:

1. Teacher-led introduction: The teacher would set the learning objectives and check students' prior knowledge.
2. Main paired activity: Students worked in pairs with one device (a 2:1 student-to-device ratio). Crucially, they took on defined roles: the "driver" typed the questions into Gemini, while the "navigator" took notes and guided the inquiry. Roles were swapped each lesson to ensure balanced participation.
3. Class discussion: The teacher led a session to consolidate what was learned.
4. Plenary summary: The teacher wrapped up with a final summary .

To prepare, teachers defined lesson objectives, wrote starter prompts, and drafted question stems on the chalkboard to scaffold how students interacted with Gemini .

The Results: A Significant Boost, But Not for Everyone Equally

The trial produced statistically significant gains, but the magnitude of the effect depended heavily on how much the tool was used and the student's starting point.

Metric	Finding
Intent-to-treat (ITT)	On average, students assigned to use Guided Learning saw a +0.258 standard deviation (SD) improvement. This is equivalent to moving a student from the 50th to the 60th percentile .
Dosage-response relationship	The 69% of students who completed the full target dosage of 12 hours or more achieved a much larger effect of +0.380 SD. Each additional hour of use was associated with a +0.016 SD gain .
Baseline proficiency	The most critical finding: students with stronger starting math skills benefitted significantly more. The treatment effect increased by 0.195 SD for every additional standard deviation of baseline math proficiency a student had .
Subgroup analysis	There were no significant differences in the AI's effect when broken down by students' baseline reading proficiency, gender, or age .

How Students Actually Interacted with Gemini

One of the most revealing parts of the study was the transcript analysis of over 113,000 student-Gemini interactions. It painted a picture of the AI acting not as an answer machine, but as a tutor.

The vast majority of conversations—91.4%—were focused on mathematical understanding . Gemini's own behavior was carefully steered by the Guided Learning feature: it posed scaffolding questions in 76.4% of its messages, pushing students to think for themselves, and provided a direct solution in just 2.1% of cases .

The Risk of Widening Attainment Gaps

The finding that initially higher-performing students gained the most prompted researchers to flag a significant concern. Without deliberate strategies to support lower-performing students, the intervention risks widening existing attainment gaps rather than closing them .

What Teachers Experienced

Qualitative feedback from teacher focus groups added another dimension to the results. Teachers reported a noticeable increase in student engagement and more active classroom participation . Many began using Gemini for lesson preparation, discovering new ways to explain complex topics like fractions .

Crucially, teachers described a shift in their own professional identity, feeling their role evolve from "lecturer" to "facilitator," moving around the classroom to support student pairs .

Not all feedback was positive. Some teachers reported challenges, including Gemini occasionally introducing off-topic content or slowing the class's progress through the planned curriculum .

The Study's Critical Limitations

While the results are positive, the researchers are careful to outline several limitations that should contextualize the findings:

• Bundled intervention: The trial combined Guided Learning with teacher training, a new lesson structure, device provision, and in-class monitoring. The design cannot isolate the effect of the AI tool alone from these other crucial supports .
• Generalizability: The sample is from a single district in Sierra Leone and is not nationally representative. The findings are most applicable to schools with comparable infrastructure, including tablets, reliable internet, and teacher readiness .

A Global Research Blueprint

The Sierra Leone trial is just the beginning. It is the first in a planned international series of preregistered trials by Google DeepMind and Fab AI to study Guided Learning across different education systems . In a move toward transparency and scalability, the partners publicly released their teacher training materials and a rapid RCT playbook alongside the results, providing a blueprint for other researchers and education systems .