AI tutors

Also known as LLM tutors, AI tutoring systems, Generative AI tutors, Chatbot tutors

AI tutors are chatbots built on large language models that help students learn through explanation, practice questions and feedback. Randomised trials report mixed results: some find gains, while others find no benefit or lower performance when students can simply ask for answers.

Pilot and early commercialDisputedReviewed 22 Oct 2026
AI illustration · Illustrative scene, not a photograph of the specific device or experiment.

In plain language

An AI tutor is a chatbot, built on a large language model, that talks with a student about a subject. It can explain an idea in several ways, set practice questions, check working and offer hints. Some tutors are general chatbots used as they are, while others add teacher-written instructions that shape how they respond.

The research so far points in different directions. Some trials found that students learned more with an AI tutor than with ordinary classroom teaching. Other trials found that students who could simply ask for answers did worse on a test taken without the tool. The results depend on how the tutor is designed, what it is compared with and how learning is measured. Most studies are short, come from a single school or university, or have not yet been peer reviewed.

Technical explanation

Most current AI tutors are instruction-tuned large language models, often GPT-4-class models, wrapped in a system prompt that sets the teaching behaviour. Design choices include whether the model gives final answers or hints first, whether it has access to worked solutions, whether it asks the student questions before responding, and whether the interface is text or voice. Bastani and colleagues (2025) compared an unrestricted GPT-4 interface with a version whose prompt included teacher-written solutions and instructions to give hints rather than full answers.

Study designs vary. Kestin and colleagues (2025) used a randomised crossover design in which each student had one lesson with a purpose-built tutor at home and one in-class active-learning lesson. Liu and colleagues (2026) compared course sections with and without integrated tutor access at a large US public university. Yang and colleagues (2026) randomised online MBA students to an AI tutor or to a holdout group. Several studies measure performance during practice, which can rise quickly, separately from unassisted tests, which can show the opposite pattern.

Two issues recur across the studies. First, gains on assisted practice do not necessarily carry over to later independent performance. Second, the comparison matters: a tutor compared with no AI, with consumer AI or with an active-learning class can give different answers. The studies report effects as standard deviations or percentage changes, and they do not agree on a single figure for learning.

What the evidence says

Each statement is typed by the strength of its evidence and linked to the source that supports it.

Demonstrated Shown in a published experiment, test or shipping product.

In a field experiment with nearly 1,000 high school maths students in Turkey, randomised by classroom, unrestricted GPT-4 access raised assisted practice scores by about 48% relative to control, but students using it scored about 17% lower on an unassisted exam.

Sources: [1] · Abstract and Results

A GPT-4 tutor whose prompt included teacher-written solutions and instructions to give hints rather than full answers produced a practice gain of about 127% and an exam result that did not differ significantly from control, so the exam harm was avoided but no exam gain appeared.

Sources: [1] · Results

In a randomised crossover trial with 194 students eligible for analysis (233 enrolled) in an introductory physics course at Harvard, students who used a purpose-built AI tutor at home had a median post-test score of 4.5, against 3.5 for in-class active learning. The AI group used a median of 49 minutes, against an assumed learning time of about 60 minutes in class. The linear model effect size was 0.63 SD, which the authors describe as an underestimate because of a ceiling effect.

Sources: [2] · Results · medium confidence

Reported, unverified Stated by a company or team but not independently confirmed.

A randomised trial of 2,379 undergraduates at a large US public university found that access to a course-integrated generative AI tutor reduced final grades by 0.37 SD and reduced learning management system participation by 0.90 SD. The authors estimate that academic losses were greater among first-generation students.

Sources: [5] · Abstract · medium confidence

A pre-registered randomised field experiment in an online MBA corporate-finance module reported that the AI tutor arm gained 6.63 points more out of 55 than holdout peers (95% CI +1.95 to +11.31). The estimate rests on 46 students after quality screens, from 86 randomised, and a covariate-light specification was not significant.

Sources: [6] · Results · low confidence

In two pre-registered laboratory experiments, Lehmann and colleagues found no overall effect of LLM access on learning. Students who used LLM output in place of their own work covered more topics but understood each one less, students who used it to support their own learning understood topics better, and LLM access widened the gap between students with low and high prior knowledge.

Sources: [3] · Abstract · medium confidence

A World Bank randomised trial in Nigeria reported that a GPT-4-based tool (Microsoft Copilot), used over six weeks, raised an overall assessment by 0.31 SD and English scores by 0.23 SD. The overall figure includes AI knowledge and digital skills, so the English effect is the better guide. Gains were largest for female students and for students with higher prior attainment.

Sources: [4] · Abstract · medium confidence

Estonia's AI Leap programme provides a ChatGPT-based learning app, designed to guide thinking rather than give ready-made answers, to all 154 upper secondary schools, about 20,000 students in grades 10 and 11. Students began receiving it in late January 2026. The programme says impact study findings will be published later in 2026, so the source gives no learning-outcome results.

Sources: [7] · Programme description · medium confidence

Disputed Contested by other credible sources.

Reported effects of AI tutors on learning are mixed. Several studies find gains, while others find no overall difference, lower grades or lower unassisted performance. The direction appears to depend on tutor design, comparison group and how learning is measured.

Sources: [1], [2], [3], [5], [6] · Synthesis across sources · medium confidence

Limitations and open problems

  • Most trials are small in scope: one high school in Turkey, an introductory physics course at one university, and a five-week online MBA module.
  • Answer-giving tools can reduce unassisted performance. In the Turkish trial, the unrestricted tool raised practice scores but lowered scores on a later exam taken without it (Bastani and colleagues, 2025).
  • Long-term effects on retention and independent skill are not known. None of the cited trials reports retention beyond its own assessment window.
  • Results depend on design, comparison group and sample. The MBA trial's main estimate was significant, but its covariate-light specification was not (Yang and colleagues, 2026). The US university trial found lower grades with tutor access (Liu and colleagues, 2026).
  • Four of the seven sources are preprints or working papers that have not been peer reviewed, as far as checked, so their figures may change.
  • The national rollout in Estonia, the AI Leap programme, has no learning-outcome results in the source. The programme says impact study findings will be published later in 2026.

Possible future uses

Speculative

Ideas, not predictions. None of these has been demonstrated.

  • Tutors that ask questions before answering, as Estonia's app is designed to do, tested for retention weeks or months after study.
  • Voice tutors for spoken language practice, a skill that the cited trials mostly did not measure directly.
  • Tutors that adapt to prior knowledge in ways that narrow, rather than widen, the gap between stronger and weaker students.

Timeline

  1. 29 Aug 2024Lehmann, Cornelius and Sting post 'AI Meets the Classroom' to arXiv, reporting lab experiments in which overall learning was unchanged but the way students used LLMs mattered. Sources: [3]
  2. 2025The World Bank publishes Policy Research Working Paper 11125, reporting a Copilot (GPT-4-powered) tutoring trial in Nigeria with a 0.23 SD gain in English. Sources: [4]
  3. Jun 2025Bastani and colleagues publish in PNAS a Turkish field experiment in which unrestricted GPT-4 access lowered unassisted exam scores and a guardrailed tutor largely avoided the harm. Sources: [1]
  4. 3 Jun 2025Kestin and colleagues publish a randomised crossover trial in Scientific Reports in which a purpose-built AI tutor produced more learning in less time than in-class active learning. Sources: [2]
  5. Jan 2026Estonia's AI Leap learning app begins reaching upper secondary students across the country. Sources: [7]
  6. Oct 2026Liu and colleagues post EdWorkingPaper 26-1598, reporting that course-integrated tutor access reduced final grades by 0.37 SD in a US university trial. Sources: [5]

Futures this could help bring about

AI illustration
How far away?

An AI tutor for every student

Pilots · 2 of 6 steps done

Glossary

Large language model
An AI model trained on large amounts of text to predict and generate language. Chatbots such as ChatGPT are built on models of this kind.
Randomised controlled trial
A study in which people or groups are assigned to conditions by chance, so that the groups can be fairly compared.
Standard deviation (SD)
A unit that expresses a score difference relative to the usual spread of scores. An effect of 0.3 SD means the average score was about three-tenths of a standard deviation higher.
Guardrail
A rule in a tutor's instructions that limits its behaviour, for example by forcing hints before full answers.

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, Jun 2025. DOI: 10.1073/pnas.2422633122. Accessed 22 Oct 2026.
  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. DOI: 10.1038/s41598-025-97652-6. Accessed 22 Oct 2026.
  3. preprint
    AI Meets the Classroom: When Do Large Language Models Harm Learning?
    Matthias Lehmann, Philipp B. Cornelius, Fabian J. Sting. arXiv, 29 Aug 2024. Accessed 22 Oct 2026.
  4. 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. Accessed 22 Oct 2026.
  5. 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, Annenberg Institute at Brown University, Oct 2026. DOI: 10.26300/y3f8-vh05. Accessed 22 Oct 2026.
  6. preprint
    When AI Tutors Speak: Evidence from a Randomized Field Experiment
    Shihao Yang, Marshall Van Alstyne, Chrysanthos Dellarocas. arXiv (working draft), 21 Sept 2026. Accessed 22 Oct 2026.
  7. 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. Accessed 22 Oct 2026.

Revision history

3 changes since this profile was created.

Changes on 22 Oct 2026
Published on owner instruction (content loop iteration 61)f3b4e10
Sources and claims re-checked against originals409337d
Initial profile drafted from primary sourceseaac9eb