An AI tutor for every student

How close are we to every student having an AI tutor that measurably improves learning?

AI tutors are in national rollout in Estonia, where an app built on ChatGPT reached all 154 upper secondary schools from January 2026. Randomised trials disagree. Some show gains on a test taken straight after study. Others show that tutors which give answers can lower later unassisted scores, and one large US university trial found lower grades with tutor access. No national learning-outcome result has been published yet, and no source opened shows a lasting gain once tutor access is removed.

AI illustration · Illustrative scene, not a photograph of the specific device or experiment.
WHERE WE AREPilots

Close. Early pilots are running in the real world.

Pilot. Tutors are in a national rollout in Estonia, which began in January 2026, and in several trials. No national learning-outcome result is published. The trials disagree on whether a tutor raises independent learning, and none opened for this record shows a gain that lasts after tutor access ends.

The tech tree

Everything recorded on Vrulum that this future needs, oldest foundations at the top and the goal at the bottom. Lines show what each step requires.

What it takes: An AI tutor for every student Unlocked Being researched Locked
  1. 2024AI tutorsPilot and early commercial Unlocked.
  1. 2025Trials report learning gains under some designsDone Unlocked.
  2. 2025Tutors that give answers can lower unassisted scoresDone Unlocked.
  3. 2026National rollout with learning-outcome evaluationIn progress Being researched.
  4. 2026Trials replicated in other courses and countriesIn progress Being researched.
  5. Date not recordedGains shown to last after tutor access endsNot started Locked.
  6. Date not recordedTutors tested across all types of school and studentNot started Locked.
  1. GoalAn AI tutor for every studentGoalRequires: AI tutors, Trials report learning gains under some designs, Tutors that give answers can lower unassisted scores, Trials replicated in other courses and countries, National rollout with learning-outcome evaluation, Gains shown to last after tutor access ends, Tutors tested across all types of school and student. Locked.

The road so far

2 of 6 steps done, 2 in progress.

  1. Done · 3 Jun 2025Trials report learning gains under some designs

    In a randomised controlled trial in an introductory physics course, 194 students were eligible for analysis. Those who used a purpose-built AI tutor had a median post-test score of 4.5, against 3.5 for in-class active learning. The linear model effect size was 0.63 SD, which the authors say is an underestimate because of a ceiling effect. A World Bank trial in Nigeria reported a 0.23 SD gain in English, but the abstract gives no sample size. [2][3]

  2. Done · 25 Jun 2025Tutors that give answers can lower unassisted scores

    In a field experiment with Turkish high school students, unrestricted access to a standard GPT chatbot ('GPT Base') raised practice scores by 48% relative to control, a gain of 0.137 on a 0 to 1 scale. On a later exam taken without the tool, those students scored 17% lower (a change of -0.054 on a 0 to 1 scale, marked P < 0.05 in the paper). A GPT tutor with teacher-written guardrails avoided that drop, but its exam result did not differ significantly from control. [1]

  3. In progress · Oct 2026Trials replicated in other courses and countries

    Two newer trials point in different directions. A preprint from a large US public university, with 2,379 undergraduates, found that course-integrated tutor access reduced final grades by 0.37 SD and reduced learning management system participation by 0.90 SD. A working draft from Boston University, with 46 students in its pre-registered analysis, found a gain of 6.63 points out of 55 with a 95% CI of +1.95 to +11.31. Without the extra controls, the estimate was not significant at the conventional 0.05 level. Both are preprints that have not been peer reviewed, as far as checked. [4][5]

  4. In progress · Jan 2026National rollout with learning-outcome evaluation

    Estonia's AI Leap programme gives its app, based on ChatGPT and designed to ask students questions before answering, to all 154 upper secondary schools, about 20,000 students in grades 10 and 11. Students began receiving it in late January 2026, after a pilot year from August 2025. As of March 2026, 7,700 students had activated accounts and 47% of activated students used the app weekly. The programme's page reports no learning outcomes, and says impact study findings will be published later in 2026. [6]

  5. Not startedGains shown to last after tutor access ends

    None of the trials opened for this record reports learning measured weeks or months after tutor use ended. The Turkish trial measured only short-term outcomes, which its authors give as a limitation. [1]

  6. Not startedTutors tested across all types of school and student

    The trials opened for this record took place in one high school in Turkey, one introductory physics course, one US public university, one online MBA module and one programme in Nigeria. Estonia's rollout covers upper secondary students only. None of the sources opened reports access or outcomes across all school types, regions or income groups. [1][2][4][5][3][6]

Why it matters

A tutor available to every student could give many learners the one-to-one help that few schools can afford. The same tools can also make students look more capable than they are while they are using them. Whether an AI tutor helps depends on how it is designed, how it is compared with other teaching and how learning is measured, so the question needs careful trials before it is answered at national scale.

What stands in the way

  • Tutors that give answers can lower later unassisted performance, so design choices need testing for independent learning as well as practice scores.
  • Trial results conflict. The US university trial found lower grades with tutor access, while the MBA trial found a gain that depended on the analysis chosen.
  • No national learning-outcome result has been published. Estonia's impact study is still pending.
  • Most trials are small or narrow, covering one school, one course, one university or one module. Results may not carry over to other subjects, ages or countries.
  • Few studies measure learning after tutor access ends, so it is not known whether any gain lasts.
  • Teachers need training and time to set tutor use within lesson plans, and the trials did not test that cost at scale.

Published forecasts

What others have predicted. These are their estimates, not ours.

WhoForecastTimeframeMade
AI Leap programme (TI-Hüpe), EstoniaImpact study findings will be published in a peer-reviewed international journal later in 2026. The forecast concerns the publication of the study, not the size of any learning effect. [6]20269 Mar 2026

Signals to watch

  • Publication of Estonia's AI Leap impact study, with learning outcomes compared against a control group.
  • Peer-reviewed publication of the US university trial and the MBA trial, with full sample sizes and sensitivity analyses.
  • Replications of the physics course tutor trial in other courses, with unassisted post-tests.
  • Retention tests taken weeks or months after tutor use stops.
  • Access and outcome data by school type, region or household income in national rollouts.

Sources

  1. peer reviewed paper
    Generative AI without guardrails can harm learning: Evidence from high school mathematics
    Hamsa Bastani, Osbert Bastani, Alp Sungu et al.. Proceedings of the National Academy of Sciences, 25 Jun 2025.
  2. peer reviewed paper
    AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting
    Greg Kestin, Kelly Miller, Anna Klales et al.. Scientific Reports, 3 Jun 2025.
  3. other
    From Chalkboards to Chatbots: Evaluating the Impact of Generative AI on Learning Outcomes in Nigeria
    Martin Elias De Simone, Federico Hernan Tiberti, Maria Rebeca Barron Rodriguez et al.. World Bank Policy Research Working Paper 11125, 2025.
  4. preprint
    The Effects of Course-Integrated AI Tutoring on Student Performance and Engagement: A Randomized University Trial
    Jing Liu, Tracy Sweet, Marty Haoyuan Chen et al.. EdWorkingPaper 26-1598, Oct 2026.
  5. preprint
    When AI Tutors Speak: Evidence from a Randomized Field Experiment
    Shihao Yang, Marshall Van Alstyne, Chrysanthos Dellarocas. arXiv (working draft), 21 Sept 2026.
  6. other
    Estonia Launches Nationwide AI Leap Education Program and Introduces Learning-Supportive AI in Upper Secondary Schools
    TI-Hüpe (AI Leap programme), 9 Mar 2026.

Last reviewed 22 Oct 2026.