The promising findings were conducted in UK classrooms, paving the way for a 2026 U.S. randomised controlled trial (RCT) with Imagine Learning.
LONDON–(BUSINESS WIRE)–One-to-one tutoring is education’s gold standard, though it has always faced an impossible scaling problem. A new exploratory study from Eedi, an evidence-based provider of AI-powered learning tools, in partnership with Google DeepMind, suggests a promising solution: a human-in-the-loop AI collaboration that measurably boosts student learning.
The trial took place in the summer of 2025 across five UK secondary school classrooms, utilizing the Eedi platform. Under the supervision of expert human tutors, we tested a model in which the core instruction was delivered by LearnLM—a family of models from Google and Google DeepMind, purpose-built for learning and grounded in pedagogical research. The first test was to confirm whether the human-AI team could match the ‘gold standard’ human tutor in real-time. The human-AI team was just as effective at helping students immediately fix a mistake they made (93.0% success vs. 91.2% for a human tutor) and at helping students resolve their underlying misconceptions (95.4% vs. 94.9%) as the expert human tutor working alone.
These new findings are particularly significant because they build on Eedi’s year-long 2023-24 randomised controlled trial (RCT) with 2,901 students in 20 UK classrooms. The treatment condition, consisting of Eedi’s static content model (high-quality diagnostic questions, videos, and hints), showed that students achieved two to four additional months of math gains over the academic year. That same high-impact intervention served as the baseline in this new exploratory trial.
The latest trial also measured “knowledge transfer”: how tutoring on one problem affected a student’s ability to solve a new problem. Compared with the standard hint, a human tutor alone improved a student’s learning by 4.5 percentage points, while the human-AI team boosted learning by 10 percentage points. The AI tutor acted as an amplifier, doubling the human tutor’s impact on this critical learning metric.
“For years, we’ve known that personalized tutoring is the most effective intervention in education, but also the least scalable,” said Dr. Bibi Groot, Chief Impact Officer at Eedi. “That’s why we started to build evidence with a small exploratory trial to learn. We found that our ‘human-in-the-loop’ AI model is safer, matches the gold standard, and shows significant promise for deeper learning. This pilot gives us the critical insights we need to design our next large-scale studies”.
This partnership also created a new, scalable model. Tutors reported the AI acted as a ‘cognitive offload,’ doing the ‘hard work’ of decoding student messages and drafting high-quality pedagogical responses. This allowed them to support multiple students at once with the focus of a one-on-one session.
The trial, conducted with 165 students and 17 tutors, was a careful test of safety and trust. A full audit of all AI-generated messages found zero harmful or risky content, and only 0.1% of messages contained any factual inaccuracies. Expert tutors, who had the final say on every message, approved 82.3% of the AI’s suggestions with no or minimal (1-2 character) edits.
The study also provides a promising blueprint for the future of tutoring. The AI was found to be a powerful, but rigid, logician. The human’s indispensable role became that of an emotional and pedagogical moderator. Tutors’ most frequent interventions were to protect the student from the AI’s blunt logic; either by managing the conversation’s pacing when its Socratic questioning ‘frustrated’ the student (44.3% of edits) or by softening its ‘transactional’ tone (19.5% of edits).
“These findings mark a milestone for responsible, safe, and effective AI in education,” said Irina Jurenka, Research Scientist at Google DeepMind. “The next step is to scale this from an exploratory pilot to a large-scale trial.”
The success of this human-in-the-loop AI model has paved the way for additional large-scale studies. Eedi is partnering with Imagine Learning to roll out a new RCT during the 2026 academic year, testing the model’s effectiveness across diverse U.S. districts. Furthermore, Eedi will conduct another UK-based RCT to assess long-term learning gains using the STAR assessment, supported by a grant funded by the Learning Engineering Virtual Institute.
“For generations, effective one-on-one tutoring has been a luxury for the privileged. Our research provides the first definitive proof that AI can safely deliver that same quality of instruction at a scale and cost that makes it universally accessible,” said Ben Caulfield, CEO and Co-Founder of Eedi. ”We are building a world where a student’s progress is no longer defined by their income, but by their potential.”
To download Eedi and Google DeepMind’s Technical Report, please visit: goo.gle/LearnLM-Nov25.
About Eedi
Eedi partners with schools, governments, and market-leading learning platforms and publishers to deliver AI-powered, evidence-based tools proven to accelerate learning. Built on more than a billion learner interactions, Eedi combines diagnostic intelligence, teacher oversight, and adaptive technology to make personalized learning measurable, scalable, and equitable. The company’s mission is to enable measurable learning progress for one billion children by 2030. Learn more at www.eedi.com.
Contacts
Alyssa Miller
Eedi
[email protected]
973-615-1292
Kristen Morea
Google DeepMind Communications
[email protected]
