AI tutors and learning outcomes Socratic and dialogic tutoring

Supervised AI tutoring matched human tutors in UK maths classrooms

In short: In five UK secondary schools, pupils who got chat tutoring drafted by Google's LearnLM model, with every message checked by an expert tutor, did at least as well as pupils tutored by humans alone, and were 5.5 percentage points more likely to solve new problems on later topics. Tutors approved most AI drafts with little or no editing.

What they did

Over seven weeks (May–June 2025), pupils were randomly assigned either to static hints (a pre-written message about the likely misconception when they got a question wrong) or to live chat tutoring. Within the tutoring group, each session was randomly given to a human tutor alone or to LearnLM, a model fine-tuned for teaching, whose every message a human tutor approved, edited or rewrote before it was sent. Pupils could not see which kind of session they had.

Who took part

165 pupils in Years 9 and 10 (ages 13–15) at five UK secondary schools using the Eedi maths platform: 91 in the static-hint group and 74 in the tutoring group.

What they found

• Supervising tutors approved 76.4% of LearnLM's drafted messages with zero or minimal edits.
• Pupils supported by LearnLM did at least as well as pupils with human tutors on every outcome measured.
• On novel problems in later topics, success was 66.2% after LearnLM-supported sessions and 60.7% after human-only sessions, a 5.5-point difference.
• A post-trial review of all 3,617 drafted messages found no harmful content and five factual errors (about 0.1%).
• Tutors said LearnLM was strong at drafting Socratic questions, and some said they picked up new techniques from it.

Effect size: +5.5 percentage points on novel problems (66.2% vs 60.7%) for LearnLM-supported versus human-only sessions.

Limitations

• Small and exploratory: 74 pupils in the tutoring arm.
• Every AI message was supervised by an expert, so this does not show how unsupervised AI tutoring performs.
• Tutors worked in both kinds of session, so what they learned from the model may have carried over and narrowed the gap.
• Short trial with platform outcomes, not standardised exams.
• Written by the teams that build LearnLM and Eedi, and a preprint at the time of summary.

What this means for learners and parents

• AI that asks guiding questions, checked by a person, can be as good as a human tutor in short maths sessions.
• "Socratic" design, finding the misconception and asking about it rather than giving the answer, was what tutors valued.
• This is evidence for supervised, purpose-built tutoring, not for any chatbot.

Put it into practice (free tools)

• How to tell whether an AI tutor asks or tells
• Tools that are Socratic in the literal sense

Try it: Ask Poppy is an AI tutor for iPhone, iPad and Mac that asks questions and gives one hint at a time instead of handing over the answer. Free with 30 messages a day, then $3.99/week or $9.99/month. (We make it.)

Citation

LearnLM Team (Google), Eedi, and colleagues (Albert Wang, Aliya Rysbek, Andrea Huber and others) (2025). AI tutoring can safely and effectively support students: An exploratory RCT in UK classrooms. arXiv preprint 2512.23633. https://doi.org/10.48550/arXiv.2512.23633

Related summaries

• Unrestricted ChatGPT help raised practice scores but lowered exam scores · Hamsa Bastani et al., 2025 · Preregistered randomized controlled trial
• A carefully designed AI tutor beat an active-learning physics class on a short-term test · Greg Kestin et al., 2025 · Randomized controlled trial with a crossover design
• An AI coach for human tutors raised topic mastery, most for weaker tutors · Rose Wang et al., 2024 · Preregistered randomized controlled trial
• Six weeks of teacher-guided AI tutoring improved English results in Nigeria · Martin De Simone et al., 2025 · Randomized controlled trial

Summary by the Ask Poppy team, AI-assisted and checked against the paper. Summarized 21 October 2026. Spotted an error? [email protected] · Editorial policy