Eedi
Eedi and Google trial LearnLM as a tutor-supervised maths helper for 165 UK students
Eedi and Google tested LearnLM, an education-tuned AI model, drafting chat messages that human tutors reviewed before students saw them, in five UK secondary schools. Tutors approved 76.4% of drafts with zero or minimal edits. Students' follow-up results were similar to those with human tutors alone, and the one apparent edge, on the next topic, was statistically uncertain.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
76.4%
Share of LearnLM's drafted messages that supervising tutors approved with zero or minimal edits (changing one or two characters), out of 3,617 drafted messages1
5
(0.1%)
Factual errors found in a post-trial review of the tutors' edits, out of 3,617 drafted messages (zero harmful or risky content found)1
66.2%
vs 60.7% (5.5 percentage points; 95% credible interval [-1.4%, +12.4%])
Students answering the first question of the next topic correctly: LearnLM-supported sessions versus human tutor alone (exploratory result)1
93.0%
vs 91.2%
Students correct on their second attempt at the same question after tutoring: LearnLM-supported versus human tutor alone1
The problem
One-to-one tutoring helps students learn, but it costs too much to give to most of them. Eedi, a UK maths learning platform used by hundreds of secondary schools, already offers pre-written hints and live chat help from remote human tutors. The question was whether a generative AI model could share the load without harming students.
What they built
Eedi and Google added LearnLM, a Google model tuned for teaching, to the chat help on the platform. When a student got a question wrong and a tutoring session started, the platform connected the student at random to a human tutor alone or to LearnLM supervised by a human tutor. LearnLM wrote each message in a step-by-step questioning style meant to help the student find their own mistake. A tutor read every message and could approve it, edit it or rewrite it before it was sent.
How they did it
The trial ran for seven weeks in May and June 2025 with 165 students aged 13 to 15 in five schools. Students were first assigned at random to static hints (91 students) or tutoring (74 students). 17 expert tutors took part, switching between talking to students directly and supervising LearnLM during the same lessons. The authors measured results with students' normal daily activity on the platform, so no extra tests were needed.
Results
Of the 3,617 messages LearnLM drafted, tutors accepted 74.4% with no edits and 76.4% with zero or minimal edits. A review by the tutors found no harmful content and five factual errors. Students helped by LearnLM answered 93.0% of second attempts correctly, against 91.2% with a human tutor alone. On the first question of the next topic, they got 66.2% right against 60.7%, a gap of 5.5 percentage points with a wide uncertainty range of -1.4% to +12.4%. The authors read this as at least as good as human tutors, with a possible edge in carrying learning to new topics. Because the range includes zero, treat the next-topic gap as unproven.
What went wrong
Tutors said LearnLM's steady Socratic style could frustrate some students, and that human tutors judge better when to push and when to ease off. About 19.5% of the tutors' edits adjusted the persona or tone LearnLM conveyed.
What to copy
If you pilot AI in a high-stakes setting, put a trained reviewer on every output, track how often they change it, and compare against your current process with random assignment.
Lessons for your project
Keep a human in the loop for anything a child sees. Here a tutor reviewed every AI message and could approve, edit or replace it, and the edit rate becomes a quality measure you can track.
Randomize at the session level so each student sees both options. Eedi used this to compare AI and human help for the same students.
Measure learning that carries over, not only fixing the current question. The only gap between AI and human help showed up on the next topic, and it was uncertain.
Audit the full corpus of AI messages after the trial. Tutors reviewed all edits and logged five factual errors and no harmful content.
Caveats
- This is an exploratory randomized trial reported as a preprint by Google's LearnLM team and Eedi, who built the system. It is not peer reviewed and is not a full production rollout.
- The next-topic gain is inconclusive and should not be read as a proven benefit. The 5.5 percentage point difference in next-topic success has a 95% credible interval of -1.4% to +12.4%, which includes no effect. The authors put a 93.6% credibility on LearnLM being better; the second-attempt gap (93.0% vs 91.2%) is small.
- LearnLM was always supervised by a human tutor who could edit or replace its messages. The results describe the combination, not the AI alone, and the paper says tutors may have picked up techniques from LearnLM that they used in their own sessions.
- The trial ran sessions at random within the same students, so it cannot show cumulative learning over time. The paper's efficiency findings (tutors handling more conversations) are anecdotal and an informal simulation.
- The paper says 74.4% of drafts were accepted with no edits at all; 76.4% is the share with zero or minimal edits. Google has since folded LearnLM's teaching abilities into Gemini and recommends Gemini 2.5 Pro for replication.
- The trial dates come from the paper's text (May through June 2025). The arXiv version is dated 29 December 2025.
Sources
- AI tutoring can safely and effectively support students: An exploratory RCT in UK classrooms
arXiv, LearnLM Team Google and Eedi · 2025-12-29 · Research paper · arxiv.org
Summarized in our own words from the sources above. Spotted an error? See corrections.