A new exploratory study by Eedi and Google DeepMind has found that a human-in-the-loop AI tutoring model significantly outperforms traditional human-only tutoring in key learning outcomes. The research, conducted in five UK secondary school classrooms during the summer of 2025, tested a system where expert human tutors worked alongside LearnLM, a family of AI models developed by Google and Google DeepMind specifically for education and grounded in pedagogical research. The results showed the human-AI team matched the performance of a human tutor alone in real-time error correction—achieving 93.0% success compared to 91.2% for a human tutor—and in addressing students’ underlying misconceptions, with 95.4% success versus 94.9%. This demonstrates that the AI-enhanced model can deliver the same high-quality, personalized instruction as the gold standard of one-on-one tutoring. The study built on Eedi’s prior 2023–24 randomised controlled trial involving 2,901 students across 20 UK classrooms. That earlier research showed Eedi’s static content model—featuring diagnostic questions, videos, and hints—led to students gaining two to four additional months of math proficiency over the academic year. This proven intervention served as the baseline for the new trial. One of the most compelling findings was in knowledge transfer—the ability to apply learning from one problem to a new, similar challenge. While a human tutor alone improved student performance by 4.5 percentage points compared to standard hints, the human-AI team boosted learning by 10 percentage points, effectively doubling the human tutor’s impact. The AI acted as a powerful amplifier, enhancing the tutor’s effectiveness. Tutors reported that the AI served as a cognitive offload, handling the labor-intensive task of interpreting student responses and drafting pedagogically sound replies. This allowed tutors to support multiple students simultaneously while maintaining the focus and personalization of one-on-one sessions. Safety and trust were rigorously tested. A full audit of all AI-generated messages revealed zero harmful or risky content, and only 0.1% contained factual inaccuracies. Expert tutors approved 82.3% of AI suggestions with little or no edits. The most common human interventions involved moderating the AI’s tone and pacing—44.3% of edits were to adjust overly rapid or frustrating Socratic questioning, and 19.5% were to soften the AI’s transactional or blunt language. The study highlights a clear division of roles: the AI excels at logical reasoning and pattern recognition, while the human tutor provides emotional intelligence, pacing, and empathy—essential for maintaining student engagement and confidence. The success of this model has paved the way for larger-scale trials. Eedi is partnering with Imagine Learning to launch a U.S.-based randomised controlled trial in 2026, testing the model across diverse school districts. A follow-up UK RCT will also assess long-term learning gains using the STAR assessment, supported by a grant from the Learning Engineering Virtual Institute. “This research provides the first definitive proof that AI can deliver the same quality of one-on-one tutoring at scale and affordability,” said Ben Caulfield, CEO and Co-Founder of Eedi. “We’re building a future where a student’s progress is determined not by their income, but by their potential.” The full technical report is available at goo.gle/LearnLM-Nov25.