Google DeepMind's Gemini math tutor adds two years of progress in eight weeks

The biggest fear with AI in classrooms has always been simple: give kids a chatbot, and they will stop thinking for themselves.

Google DeepMind just dropped some of the most rigorous field evidence to date showing the exact opposite can happen.

DeepMind published findings from an eight-week pre-registered randomized controlled trial (RCT) in Sierra Leone's Port Loko District. Run alongside Fab AI, the Gates Foundation, and the local Ministry of Education across 12 schools, the project logged over 113,000 interactions with a specialized Gemini math tutor.

The headline numbers are eye-popping. Students using Gemini's Guided Learning posted a test score gain of +0.258 standard deviations, representing roughly 1.2 to 1.7 years of typical math progress crammed into just eight weeks. In classrooms where teachers leaned in heavily—hitting around 12 hours of integrated instruction—gains skyrocketed to 1.8 to 2.5 years.

The tutor that refuses to do your homework

Gemini was tuned using Google's pedagogical LearnLM research, basically turning it into a stubborn Socratic coach.

Instead of acting like an answer key, the AI asked scaffolding questions in 76% of its responses. It handed over direct answers only 2% of the time, forcing students to do the cognitive heavy lifting themselves.

More impressive than the test scores was the behavioral flip. Educational software notoriously suffers from the "5% problem," where barely five percent of users stick around voluntarily. Here, 69% of students hit or exceeded their target usage.

Over the trial, pure solution-seeking prompts plunged from 25% to 10%. At the same time, skill-building queries climbed from 68% in week one to 90% by week eight. Fully 91.4% of student conversations focused on building conceptual understanding rather than hunting for shortcuts.

Teachers at the wheel

Despite techno-utopian dreams of replacing human instructors, the trial proved that AI works best as an amplifier for teachers, not a substitute.

Educators used Gemini to prep lesson plans, finding fresh ways to explain tricky topics like fractions. Instead of lecturing from the blackboard, teachers moved around the room acting as facilitators for student pairs.

To help other regions replicate the model, DeepMind released a practical teacher training guide and an RCT research playbook.

There is one major red flag in the data, though: the achievement gap. Students who started the trial with stronger math skills saw the largest gains, highlighting a persistent risk that AI could widen disparities if left unchecked.

Why it matters

While Silicon Valley has spent months hyping AI tutors with thin anecdotal evidence, this pre-registered RCT gives us hard, peer-ready data. It proves that prompt design and pedagogical guardrails genuinely matter: when you build an AI that refuses to spoon-feed answers, kids actually rise to the challenge.

We think DeepMind's smartest move was tethering the AI directly to classroom teachers rather than pitching a standalone self-study app. The fact that learning gains nearly doubled when teachers were actively involved should kill off the fantasy of teacherless AI classrooms once and for all.

For ed-tech founders and school systems, the study provides a playbook worth studying right now. But we should also take the headline gains with a grain of salt. In resource-constrained settings like Sierra Leone, novel technology and structured attention produce outsized honeymoon effects.

Whether this Socratic approach delivers the same 2-year leaps at national scale—and whether it can lift struggling students without favoring top performers—remains the real hurdle ahead.