Google's AI Tutor Delivered Year-Long Learning Gains in 8 Weeks

Google rebuilt Gemini to teach, not just assist

Google DeepMind has published results from a Sierra Leone classroom trial showing its AI tutor helped students achieve more than a year's worth of math learning in just eight weeks. The tool, called Guided Learning, represents a fundamental rebuild of Gemini specifically designed for pedagogy rather than assistance.

Irina Jurenka, the research director who led the project, cautions that the "year of progress" figure is an estimate extrapolated from a measured 0.26 standard deviation improvement. The translation to calendar time relies on limited literature, mostly from literacy studies rather than mathematics. Still, the underlying gains are real and appear to generalize beyond the specific material taught.

Why it matters

Most AI assistants are built to complete tasks for users—the opposite of what learning requires. This trial demonstrates that purpose-built AI can shift student behavior away from answer-seeking and toward genuine problem-solving, with measurable academic outcomes. For education leaders evaluating AI pilots, the distinction between generic chatbots and pedagogically designed tools now has empirical support.

The technology behind the pedagogy

Jurenka's team started where most schools do: writing prompts that tell a language model to act like a teacher. That approach failed. "It's kind of the same as writing a one-pager on teaching, handing it to an average stranger on the street and expecting them to come into a classroom and teach," she explained.

The solution required technical work at the model level. In the classroom version of Guided Learning, the system asked guiding questions in 76% of its messages and provided direct answers just 2% of the time. When students repeatedly asked for answers—a common tactic with AI assistants—they eventually stopped nagging and engaged with the learning process instead.

Two versions of Guided Learning exist. The public consumer app respects user agency and will eventually provide answers if pressed. The teacher-led classroom version, used in Sierra Leone, never gives answers away.

Stronger students gained most

The trial revealed an equity concern: students already performing well in mathematics saw the largest gains. Jurenka acknowledged this immediately and said DeepMind is now partnering with organizations specializing in supporting struggling learners to explore different pedagogical approaches.

A separate World Bank and Stanford trial in Nigeria found similar-sized gains but showed girls who started behind improving more, suggesting design choices matter more than the underlying model.

Teachers adapted their practice

The intervention required only one day of teacher training and was designed with local educators. Teachers reported that Gemini sometimes explained concepts differently than they would, which they found valuable. The technology enabled them to spend more time moving between students for one-on-one conversations rather than delivering whole-class instruction.

Usage rates far exceeded typical voluntary education technology. While most edtech sees roughly 5% sustained engagement, 69% of students in this trial met or exceeded their usage targets.

Transparency and replication

DeepMind published its complete methodology, including a playbook documenting trial design, measurement protocols, and teacher training materials. The team separated decision-making from measurement to minimize bias and explicitly invited replication studies.

"Our research is not about creating proof points for Google," Jurenka said. "It's about understanding the broader picture of where AI fits in, in improving learning."

The independent end-of-term assessment was designed by external assessors and included 50% material covered during the term plus 50% broader mathematical topics to test for generalization. Students completed the assessment on paper, without AI assistance.

Details of the Sierra Leone trial were first reported by Dan Fitzpatrick in Forbes.