Study: AI tutors like LearnLM provide high-quality, individualized support in real classrooms
As reported in The 74, a new study analyzes a real-world classroom trial of LearnLM, an AI “math tutor” plugged into the UK platform Eedi and used with 13–15-year-old students.
For the study, students were randomly split into groups. Some were given traditional, pre-written hints. Others were provided live chat support, either from a human tutor, or from LearnLM. LearnLM messages were always checked and minimally edited (only as needed) by expert tutors before students received them.
Across more than 3,600 AI-generated messages, tutors readily approved or barely edited about 75% of them. A subsequent review found almost no factual mistakes and no harmful content. In other words, the AI could provide safe, mathematically sound help without much amendment.
In terms of learning, students working with the AI tutor did at least as well as those with a human tutor, and were slightly more likely to solve new, transfer-style problems correctly. Tutors especially appreciated how the AI modeled good questioning, often using a Socratic style that pushed students’ thinking. Humans still added important touches, such as tone, encouragement, and pacing.
Tom’s Take: This is an early, small-scale study, and clearly branded by Google. But Google’s DeepMind had already proven its abilities at the International Math Olympiad and Gemini 3, its newest model, has surpassed prominent math benchmarks. High-performing math AI models don’t necessarily translate into outstanding tutors, but, in all, the study suggests that well-designed, supervised AI tutors can realistically help scale individualized math support for more students.
See also:
Related posts from Tom Daccord’s free newsletter:
From Benchmarks to Classrooms: What GPT-5.1 and Gemini 3 Really Mean for Math Teaching
The NotebookLM Learning Loop: Turning AI Into Intellectual Friction (Not an Answer Machine)
NotebookLM for Ages 13 and Up: Guiding Students Into Inquiry-Based Learning
NotebookLM in Action: Takeaways from Testing Google's AI Research Assistant
Also, visit:
