Loading article…
Google DeepMind's LearnLM AI tutor achieved 93% mistake‑fix success and doubled knowledge‑transfer impact in a 165‑student UK pilot, signaling scalable
A pilot of 165 students in five UK secondary classrooms showed that a human‑in‑the‑loop AI tutor built on Google DeepMind’s LearnLM model matched the “gold‑standard” human tutor on immediate error correction (93.0% vs 91.2%) and on fixing underlying misconceptions (95.4% vs 94.9%) [1]. The result suggests AI can safely replicate one‑to‑one tutoring effectiveness at scale, a breakthrough for education technology.
| At a glance | |
|---|---|
| Trial size | 165 students, 17 tutors |
| Immediate error‑fix success | 93.0% (AI‑human team) vs 91.2% (human only) |
| Misconception‑resolution | 95.4% vs 94.9% |
| Knowledge‑transfer boost | +10 pp vs +4.5 pp for human tutor |
| Safety audit | 0.1% factual errors, 0 harmful content |
The exploratory study, run in summer 2025, paired expert tutors with LearnLM, a family of models purpose‑built for learning and grounded in pedagogical research. Tutors acted as moderators, approving 82.3% of AI‑generated suggestions with only minor edits, while the AI handled the “hard work” of decoding student messages and drafting responses. This “cognitive offload” let tutors support multiple learners simultaneously, a key advantage over traditional one‑to‑one tutoring that is limited by human capacity.
A notable metric was knowledge transfer: when students tackled a new problem after tutoring, the AI‑human team added 10 percentage points to learning gains, double the 4.5‑point lift achieved by a human tutor alone. The AI therefore functioned as an amplifier, extending the impact of the human tutor on deeper learning outcomes [1].
LearnLM’s performance builds on earlier internal claims that the model outperforms GPT‑4o, Claude 3.5 and Google’s own Gemini 1.5 on adherence to learning‑science principles [2]. While those claims are company‑reported, the UK pilot provides independent evidence that a human‑AI team can meet or exceed human‑only tutoring on core metrics. Prior Eedi research showed static content interventions yielded two‑to‑four months of math gains over a year [1]; the new AI‑augmented approach promises comparable or greater gains without the same scaling constraints.
The trial also demonstrated robust safety: a full audit found zero harmful content and only 0.1% factual inaccuracies, addressing common concerns about AI‑driven instruction. Tutors’ edits focused on softening the AI’s “transactional” tone (19.5% of edits) and managing pacing when Socratic questioning frustrated students (44.3% of edits), highlighting the complementary role of human oversight [1].
Buoyed by the pilot, Eedi will partner with Imagine Learning to launch a randomized controlled trial across diverse U.S. districts in the 2026 academic year, extending the test to a larger, more varied student population [1]. A second UK‑based RCT will assess long‑term gains using the STAR assessment, funded by the Learning Engineering Virtual Institute [1].
The study marks a milestone: AI can now safely deliver tutoring quality comparable to human experts, potentially democratizing access to personalized education that has traditionally been limited to higher‑income students. The upcoming large‑scale trials will determine whether this promise holds in broader, real‑world settings.
Coverage is mostly measured — 132 of 144 reports stay neutral.
Every Monday — the token unlocks, Fed dates & catalysts set to move crypto and markets this week. So you’re never blindsided.
Free · 3-min read · one-click unsubscribe
AI-assisted synthesis by the TrendWatcher Editorial Desk · sourced from 2 outlets · Jul 7, 2026 · How we report
The trial claimed that students achieved a gain equivalent to roughly one year of additional schooling over an eight‑week period.
In the trial, the model provided direct answers in only about 2% of its messages, focusing instead on asking guiding questions.
Yes, the endline assessment included 50% material covered during the term and 50% broader mathematical topics to test generalization.
The trial showed larger gains for already‑strong math students, prompting researchers to consider how to adapt prompts to better support lower‑performing learners.
The Sierra Leone trial used the consumer version of Guided Learning under teacher supervision; a separate teacher‑led version with stricter answer restrictions has since been announced.