Google AI Tutor Shows Strong Learning Gains in Classroom Trial

Google DeepMind has released findings from a classroom study in Sierra Leone suggesting its AI-powered tutoring system, Guided Learning, significantly improved student learning in just over eight weeks. The project, led by DeepMind research director and computational neuroscientist Irina Jurenka, reported an average improvement of 0.26 standard deviations, which researchers estimate is roughly equivalent to more than a year of additional schooling. Jurenka emphasized that this conversion is an approximation based on existing education research and should be interpreted cautiously.

Unlike conventional AI assistants that often provide direct answers, Guided Learning was specifically redesigned from Google’s Gemini model to encourage students to solve problems independently. Jurenka explained that simply instructing a language model to behave like a teacher proved insufficient, leading the team to redesign the system around established teaching methods. During the Sierra Leone trial, the AI responded with guiding questions in 76% of its interactions, while providing direct answers only 2% of the time.

Google created two versions of Guided Learning. The publicly available version respects user choice and may eventually provide answers if repeatedly requested, while the classroom edition used in the study was configured to avoid revealing solutions altogether. According to Jurenka, students gradually stopped repeatedly asking for answers and instead became more engaged with the learning process.

To measure learning fairly, the final assessment was created by an independent evaluator and completed with pen and paper without AI assistance. Half of the exam covered material taught during the program, while the remainder tested broader mathematical understanding. Researchers said the results suggested students developed knowledge that extended beyond the lessons delivered during the trial. Student participation was also notably high, with 69% meeting or exceeding usage goals compared with the much lower engagement often seen in voluntary educational technology programs.

The study also found that students who already had stronger mathematics skills experienced the greatest improvements. Jurenka acknowledged the concern that AI could widen achievement gaps and said Google DeepMind has already begun working with education partners to explore teaching methods that better support students who are further behind. She added that different instructional approaches and prompts may be needed to help raise overall learning outcomes.

Google noted that similar improvements were observed in a separate study conducted by the World Bank and Stanford in Nigeria, although that research found girls who initially lagged behind benefited the most. According to Jurenka, these differing outcomes suggest that educational design and teaching strategies, rather than the AI model itself, play a key role in determining who benefits most.

Addressing concerns about research conducted by Google using its own technology, Jurenka said the team’s objective was to understand AI’s role in education rather than promote Google products. To encourage transparency, DeepMind published its research methodology, teacher training resources and evaluation framework so that other researchers can replicate or expand on the study while minimizing potential bias.

The project also focused on supporting teachers instead of replacing them. Educators received about one day of training and reported that the AI introduced new ways to explain concepts while allowing them to spend more time working individually with students. Jurenka said the experience suggests AI could gradually become part of classroom instruction through collaboration between educators and technology, although she cautioned that the current findings cannot yet be assumed to apply beyond the specific conditions of the Sierra Leone trial.