The studies at a glance
| Study | Who and how long | What it found | Main limit |
|---|---|---|---|
| Harvard physics (2025) | 194 university students; one lesson each, AI tutor against an active-learning lesson | Higher post-test scores with the AI tutor: median 4.5 against 3.5 | One elite university, one short lesson; the tutor relied on expert-written solutions |
| Turkey high school (2025) | About 1,000 students, four 90-minute math sessions, three groups | Both AI groups did better in practice. On the exam without AI, the unguarded chatbot group did worse than no AI; the tutor-style group did about the same as no AI | Short term, one school; the safeguarded tutor did not beat no AI on the exam |
| Nigeria after-school pilot (2025) | About 759 students analyzed, six weeks, 12 sessions | +0.31 SD overall and +0.23 SD on English | Volunteers, uneven drop-out between groups, teacher supervision |
| Ghana math tutor (2024 preprint) | About 1,000 students in 11 schools, eight months | 0.37 SD on math growth | 11 schools assigned as groups; not peer reviewed |
| UK secondary math (2025 preprint) | 165 students, human tutors supervising every AI message | At least as well as human tutors on later problems | Small, exploratory; the AI was supervised |
| GCSE science (2026 preprint) | 644 students completed, four weeks | g = 0.33 against self-directed revision | 30.7% drop-out; control was unsupervised revision; the authors call it preliminary |
| Tutor CoPilot (2024 preprint) | Hundreds of human tutors given an AI assistant | About +4 percentage points on topic mastery, about +9 for students of lower-rated tutors | It helps human tutors, not students directly; the outcome is a short quiz at the end of a session |
The benchmark: human tutoring
Tutoring by people is among the best-supported interventions in education. A 2024 meta-analysis of randomized trials of tutoring for school students found an average gain of about 0.29 of a standard deviation. A larger 2024 analysis found 0.42 overall but only about 0.16 to 0.21 for programs serving hundreds of students, so effects shrink at scale. The Education Endowment Foundation in England estimates about five months of extra progress, with the best results from short sessions several times a week for up to ten weeks, linked to what is taught in lessons.
The famous “two sigma” result from 1984 came from short experiments in which tutoring was bundled with extra quizzing and retesting. Later and larger studies find much smaller effects.
What the AI studies show, and do not
Two experiments are the best evidence so far on AI tutoring, and they point in different directions depending on how the tool was built.
A carefully built AI tutor can teach well
In the Harvard trial each student had one physics lesson from a custom AI tutor and one from an in-class active-learning lesson. The AI-tutor lesson produced higher scores on a quiz taken right after. The authors are clear about the limits: one university, one subject, one short lesson, and a tutor that depended on expert-written, question-specific solutions.
Design decides whether it helps or harms
In the Turkish study students practiced with a plain chatbot, with a tutor-style version that used teacher-designed hints and was told not to give answers away, or with no AI. During practice the tutor-style group did best of all. On the exam, without AI, the plain-chatbot group scored below the no-AI group, and the tutor-style group scored about the same as no AI. The lesson is not that AI raises exam scores; it is that an AI which does the work for the student can undermine learning, and that guardrails largely remove the harm.
Results in other settings are promising and early
The Nigeria pilot, the Ghana study, the UK math trial and the GCSE science trial all point in a positive direction. Three are preprints, two have notable drop-out, and each used a different kind of AI and a different comparison, so none can be read as a general verdict.
What we do not know yet
- Long-term effects: almost every study measures learning within days or weeks.
- Effects at scale and with ordinary teenagers at home, outside a trial.
- Which design choices matter most: hints, marking, memory, voice, length of session.
- How AI tutoring compares with an equally priced human tutor for the same hours.
What to look for in an AI tutor, given the evidence
- It asks before it tells. A tutor that does the thinking for you looks helpful and can harm.
- It reads your work and explains the mistake. Feedback on what you wrote is one of the better-supported parts of how people learn.
- Regular, short sessions. Several short classes a week beat one long one.
- Retrieval practice. Answering questions from memory, spaced over days, is among the best-supported study techniques.
- It is honest about its limits. It tells you when it is unsure and it does not replace your teacher.
Where Tutix stands
Tutix has not published a learning-outcome study of its own, so we cannot tell you it raises scores. We built Tutix with the findings above in mind: the tutor asks for your answer and reads it, homework hints are built to point the way rather than give the answer away, ready-made programs use lessons and questions reviewed in advance, and you choose how long and how often classes run. We have not tested whether that works better, so if outcomes matter to you, test it: take a class on a topic you find hard, then try a similar question a week later with no help. See how a class works.
Questions people ask
Does AI tutoring work?
Early evidence is promising but thin. A Harvard trial found an AI tutor taught more physics than an in-class active-learning lesson, and a World Bank pilot found English gains, but a school experiment found an unguarded chatbot lowered later exam scores. Design matters.
Is AI tutoring as good as human tutoring?
Nobody knows yet. Human tutoring has a large evidence base, averaging about 0.29 of a standard deviation in randomized trials, smaller at scale. AI tutoring has a handful of short or small studies.
Can a chatbot harm learning?
Yes, if it does the work for the student. In a Turkish high school experiment, students given an unguarded chatbot scored lower on a later exam, without AI, than students who never had it. A tutor-style version with safeguards avoided that harm.
What makes an AI tutor effective?
The research points to asking before telling, giving feedback on and explaining the student's own work, short regular sessions and practice from memory. How the tool is designed matters more than the fact that it uses AI.
Has Tutix been tested in a study?
Tutix has not published a learning-outcome study. Tutix was built with the design findings of the research in mind, and you can test it yourself with a topic you find hard and a similar question a week later.
Sources
- AI tutoring outperforms in-class active learning, Scientific Reports (June 3, 2025)
- Generative AI without guardrails can harm learning, PNAS (June 25, 2025)
- From Chalkboards to Chatbots (Policy Research Working Paper 11125), World Bank (May 2025)
- Effective and Scalable Math Support: Evidence on the Impact of an AI-Tutor on Math Achievement in Ghana, arXiv (preprint) (February 2024, revised May 2024)
- AI tutoring can safely and effectively support students: An exploratory RCT in UK classrooms, arXiv (preprint) (December 29, 2025)
- Evaluating AI Tutoring at the Speed of Innovation: Practitioner-Led Micro-Randomised Trials of an AI Tutoring Platform in GCSE Science, arXiv (preprint) (September 13, 2026)
- Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise, arXiv (preprint) (October 2024, revised since)
- The Promise of Tutoring for PreK-12 Learning: A Systematic Review and Meta-Analysis of the Experimental Evidence, American Educational Research Journal (2024)
- What Impacts Should We Expect from Tutoring at Scale?, Annenberg Institute, Brown University (EdWorkingPaper) (October 2024)
- Teaching and Learning Toolkit: one to one tuition, Education Endowment Foundation (review last updated July 2021)
- Improving Students' Learning With Effective Learning Techniques, Psychological Science in the Public Interest (2013)
- Test-enhanced learning: taking memory tests improves long-term retention, Psychological Science (2006)
Keep reading
GCSE qualifications are regulated in England and awarded by exam boards such as AQA, Pearson Edexcel and OCR, none of which is affiliated with, or endorses, this site.
