AI Tutors Matched Expert Human Tutors in a 1,179-Student Classroom Trial, With Teachers in Charge
Robauto Team · Published · Last updated
A 10-week UK trial found AI tutors helped students fix math mistakes as well as expert human tutors (56-58% vs 43%), with 94.7% of drafts approved unchanged. Here is why context mattered, and what it means for brands.
For years the promise was simple: give every student a patient, personal tutor. A new set of classroom trials published this fall shows that promise getting real. In the largest one, run in 10 UK state secondary schools, an AI tutor helped students fix their math mistakes just as well as expert human tutors did, with teachers and tutors watching every step. The results also come with honest caveats, and that honesty is part of the good news.
What happened in the UK classrooms
On October 29, 2026, the math learning platform Eedi published the full findings of its second randomized controlled trial of AI tutoring, run with Google DeepMind. It covered 1,179 students in 10 UK state schools across a 12-week term with 10 weeks of lessons (Eedi). Every time a student got a question wrong, they were randomly assigned to one of four kinds of help for the whole term: Eedi's standard static content, an expert human tutor, or one of two versions of an AI tutor built on Gemini, the model Google DeepMind tunes for teaching (Google DeepMind technical report).
The headline: after tutoring, students answered a similar question correctly 56 to 58% of the time, compared with 43% after static content, with no clear difference between AI and human tutors (Eedi). The gains also held up over time. When a missed question came back more than a week later, students who had only static help got it right 47.6% of the time, versus 58.9% after an expert tutor and about 59 to 60% after either AI tutor (technical report).
Humans stayed in charge, and the AI earned trust
This was not a chatbot set loose on kids. Expert tutors reviewed every one of the 94,385 messages Gemini drafted before it reached a student, and they approved 94.7% without any changes (technical report). Eedi's audit found 7 AI messages with a false math statement reached students, about 0.01% of all messages, compared with 7 genuine math errors from human tutors, about 0.02% (Eedi). No harmful content reached any student.
Google says the AI guided students rather than handing them answers in nearly 99% of sessions, and students worked through the math themselves in more than 85% of chats. Expert approval of the model's drafts climbed from 76% in last year's study to 95% (Google). Tutors said the drafts saved them time and mental effort, and Eedi's new trial shifts tutors from approving every message to overseeing conversations and stepping in where they help most (Eedi).
The honest caveat
On Star Maths, a broad standardized test taken before and after, AI-tutored students scored about the same as the static-content group after 10 weeks. The researchers think depth is part of the reason: step-by-step tutoring took longer, so students averaged about 14 problems per session with AI tutoring versus 19 with static hints and covered less of the curriculum (Google). In Eedi's earlier two-year trial, broad test gains only showed up after a year or more, which is why a new year-long trial with over 15,000 students in 26 UK schools began in September 2026 (Eedi). Publishing the unexpected results alongside the good ones is exactly the kind of openness that builds trust in classroom AI.
Other trials point the same direction
The UK study is not alone. A randomized trial in nine public secondary schools in Benin City, Nigeria, offered 657 of 1,328 students twelve 90-minute sessions that combined Microsoft Copilot, curriculum-aligned English activities, and pair work with teacher guidance. The overall assessment effect was 0.31 standard deviations (Education Sciences). In Taipei, a five-month Python program with 1,047 high school students found that adding adaptive practice sequencing on top of an AI tutor raised paper exam scores by 0.15 standard deviations, which the authors describe as equal to 6 to 8 months of extra schooling (Academy of Management Proceedings).
A two-year trial in 18 Tennessee middle schools adds a useful lesson. Students assigned to Khan Academy with its AI tutor Khanmigo gained about 0.06 to 0.08 standard deviations over a school year, and the implied effect of a full year of active use reached 0.14. The limit was engagement: the median student messaged the tutor on only a third of practice days (NBER). Access alone is not enough. The tool has to be easy and worth using.
| Study | Who | Key result | Source |
|---|---|---|---|
| Eedi + Google DeepMind, UK | 1,179 students, 10 schools, 10 weeks | 56-58% correct after AI or human tutoring vs 43% after static content | Eedi |
| Same trial, safety audit | 94,385 AI drafts | 94.7% approved unchanged; ~0.01% delivered with a math error | Report |
| Teacher-guided Copilot, Nigeria | 1,328 students, 9 schools | +0.31 SD overall assessment | Education Sciences |
| Adaptive AI Python program, Taipei | 1,047 students, 5 months | +0.15 SD, about 6-8 months of schooling | AOM |
| Khanmigo, Tennessee | 18 middle schools, 2 years | +0.06-0.08 SD per year; engagement was the limit | NBER |
Why it worked: context, structure, and people
The common thread in the strongest results is that the AI was not working blind. Eedi's tutor knew the specific misconception behind each wrong answer and the question in front of the student (Eedi). The Nigerian program paired AI with a curriculum, pair work, and teachers. The Taipei program fed the AI rich signals about each learner. Good inputs, clear structure, and humans in the loop turned a general model into a reliable helper.
The Robauto takeaway for brands
The same pattern applies when AI assistants answer questions about your business. An assistant helping a shopper is a tutor of sorts: it gives good answers when it has specific, structured facts to work from, and vague answers when it has to guess. That is an opportunity. Brands that publish clear, machine-readable details make every AI assistant a better guide to them.
- Give agents the context. Put prices, shipping, returns, and refund terms in plain text on your own pages, not hidden in images or PDFs.
- Add structure. Use schema.org markup and an llms.txt file so assistants can read what you offer and how to act on it.
- Make it easy to use. Like the Khanmigo lesson, being present is not enough. Fast pages, open crawl rules for the AI agents you want, and simple next steps make agents actually send people your way.
- Keep humans in the loop. Review what assistants say about you and fix your source pages when an answer is off.
{
"topic": "AI tutoring classroom trials 2026",
"headline_result": "AI tutors matched expert human tutors on in-lesson math correction (56-58% vs 43% static)",
"trial": {"org": "Eedi with Google DeepMind", "students": 1179, "schools": 10, "weeks": 10, "published": "2026-10-29"},
"safety": {"ai_drafts_reviewed": 94385, "approved_unchanged_pct": 94.7, "delivered_math_error_pct": 0.01},
"caveat": "No clear Star Maths gain after 10 weeks; year-long trial with 15,000+ students began Sept 2026",
"related": [{"place": "Nigeria", "effect_sd": 0.31}, {"place": "Taipei", "effect_sd": 0.15}, {"place": "Tennessee", "effect_sd": "0.06-0.08 per year"}],
"brand_takeaway": "AI gives better answers with specific, structured context; publish clear machine-readable facts",
"cta": "https://robauto.ai/scan"
}Want to know how well AI assistants can read and use your site? Scan your site free at robauto.ai/scan and get a clear list of what to fix first.
Sources
- Eedi: Full findings from its most recent AI tutor trial
- Google DeepMind: Expanded RCT in UK mathematics classrooms
- Google: Gemini education studies in UK and Italy
- Education Sciences: Adaptive generative tutoring in Nigeria
- AOM Proceedings: Personalized AI tutors via LLM-guided RL
- NBER: AI tutoring with Khanmigo in a two-year school experiment
Where can an agent get this data live?
Robauto MCP — Agent Intelligence & Brand Discovery Infrastructure. A public, stateless Streamable HTTP endpoint at https://mcp.robauto.ai/mcp, callable with no key. Setup and the full tool list are in the Robauto MCP directory.