A Gemini tutor raised math scores in a Sierra Leone trial

Google DeepMind ran a randomized controlled trial of an AI tutor in Sierra Leone, and the results are unusually concrete for a field full of optimistic pilots. Working with Fab AI and the country's Ministry of Education, the team gave 1,763 junior secondary students across 12 schools in Port Loko District access to Guided Learning, a feature in Gemini built on the LearnLM pedagogy research. Over eight weeks, students using the tool gained 0.258 standard deviations in math over a control group, which the researchers translate to roughly 1.2 to 1.7 years of typical learning progress. In classrooms where teachers integrated it more fully, the gain reached the equivalent of 1.8 to 2.5 years.

What makes the study worth reading is the interaction data behind those scores. The team analyzed more than 113,000 student exchanges with the model. In 91.4% of conversations students were working to understand a concept rather than asking for the answer, and the model pushed in that direction: it responded with scaffolding questions in 76% of its messages and handed over a direct solution only 2% of the time. Adoption was high too, with 69% of students meeting or exceeding usage targets, far above the single-digit rates common for voluntary education software.

The honest caveat is in the report: students who already had stronger math benefited most, so the tool can widen the gap it is meant to close. You can read the full writeup on the DeepMind blog.

Why it matters

If you build or fund education technology, this is a rare piece of field evidence that a general-purpose model, steered toward asking instead of answering, can move real test scores. The catch is the equity gap, so measure who benefits before you scale.