Gemini classroom trial lifts problem-solving queries from 68% to 90%

From answer seeking to problem solving

Google DeepMind’s latest education post puts a concrete number on a central AI-in-schools question. In an eight-week classroom trial, Gemini queries about how to approach problems rose from 68% to 90%. DeepMind wrote on X that students were using AI to understand concepts, not just collect answers. The source tweet is available here.

The linked research describes a randomized controlled trial in Sierra Leone’s Port Loko District, run with Fab AI and support from the country’s education ministry. It involved 1,763 junior secondary students across 12 schools and focused on Guided Learning in Gemini for math. DeepMind says the study analyzed more than 113,000 interactions, with 91.4% of student conversations classified as building conceptual understanding rather than simply asking for solutions.

That distinction matters because the strongest criticism of classroom chatbots is not that they are useless, but that they may let students bypass the work of thinking. DeepMind’s account says Gemini responded with scaffolding questions in 76% of its messages and gave direct solutions in only 2% of cases. In other words, the product design tried to keep the cognitive work with the student while giving teachers another way to extend guided practice.

Google DeepMind usually uses its account for model, research, and responsibility updates, so this tweet fits a broader push to measure AI outcomes beyond demos and test scores. The next test is replication. Eight weeks, 12 schools, and one district are meaningful, but policy decisions will require longer follow-up across languages, subjects, and infrastructure conditions. The number to watch is whether the 68%-to-90% shift survives outside a carefully supported trial.