Eedi × Google DeepMind LearnLM — Human-in-the-Loop AI Tutoring RCT in UK Secondary Schools

In summer 2025, Eedi and Google DeepMind ran an exploratory randomised controlled trial testing whether LearnLM, a generative AI model fine-tuned for pedagogy, could support human maths tutors without compromising instructional quality. The trial involved 165 students across five UK secondary schools on Eedi's chat-based maths tutoring platform, building on Eedi's earlier 2023-24 RCT of 2,901 students across 20 UK classrooms.

The study set out to test whether AI-drafted tutoring messages, reviewed and editable by an expert human tutor before reaching the student, could match or improve on human-only tutoring, while keeping a human reviewer in control of every message sent.

In the human-AI condition, LearnLM drafted tutoring messages that an expert human tutor reviewed and could edit before they reached the student. Supervising tutors approved 76.4% of LearnLM's drafts with zero or only minor edits, and some AI-drafted Socratic questions reportedly taught tutors new pedagogical techniques.

The human-AI team helped students fix an immediate mistake about as effectively as the human tutor alone (93.0% vs 91.2%) and resolve underlying misconceptions similarly well (95.4% vs 94.9%). The clearest advantage appeared on transfer: compared with a standard hint, human-only tutoring improved learning by 4.5 percentage points, while the human-AI team improved it by 10 percentage points, and students supported by LearnLM were more likely to solve novel problems on subsequent topics (66.2% vs 60.7%).

The authors describe the trial as exploratory and call for larger studies before broader conclusions are drawn: the sample of 165 students across 5 schools is small relative to the claims, and no equity-disaggregated results were reported.