AI Tutors for Kids: What the Research Shows in 2026

An AI tutor is software that explains, asks questions and gives feedback to one learner at a time. The evidence is promising but narrow: the strongest trials involve high-school and university students, results depend on how the tutor is designed, and a plain chatbot can even harm learning. No trial of AI tutoring with preschoolers was found, so for young children, a person is still the tutor.

In short

• Human tutoring works: a 2024 meta-analysis of randomized pre-K–12 trials found an average effect of 0.29 standard deviations, largest in earlier grades.
• Bloom's famous "2 sigma" came from small 1980s studies that bundled several interventions, and it has never been replicated.
• The biggest AI-tutor result so far, more than double the learning gains, was with Harvard undergraduates over two physics lessons.
• In a trial with nearly 1,000 high-school students in Türkiye, a plain GPT-4 tutor lowered later exam scores by 17%; a version with safeguards largely avoided that harm.
• Choose tutors that ask rather than tell, stay involved, and check that your child can do the work without it.

What is an AI tutor, and how is it different from a chatbot?

An AI tutor is built to teach; a general chatbot is built to answer. The OECD's Digital Education Outlook 2026 says systems that combine generative AI with "explicit pedagogical models, such as structured tutoring strategies", "show more promise than general‑purpose chatbots", while noting the evidence "is still emerging".

Tutoring software is not new. A meta-analysis of 50 evaluations of older, pre-generative intelligent tutoring systems (Kulik and Fletcher, 2016) found a median gain of 0.66 standard deviations, with much larger gains on locally made tests than on standardized ones. Read big numbers carefully.

An AI tutor is still screen time, so our guide to screen time and digital balance applies here too.

How well does tutoring work? The truth about the "2 sigma"

Tutoring works well, but not as miraculously as the legend says. In 1984, Benjamin Bloom reported studies in which pupils in grades 4, 5 and 8 were tutored in probability and cartography. The average tutored student scored about two standard deviations above classes of about 30, "above 98% of the students in the control class."

A 2024 re-examination in Education Next (von Hippel) says: "It wasn't just tutoring, and it's never been replicated." Those studies mixed tutoring with teacher coaching, extra time and frequent testing. About one-third of a standard deviation "seems to be the typical effect of an intense, well-designed program."

Larger reviews agree:

• VanLehn (2011) found human tutoring at d = 0.79, and step-based tutoring systems nearly as effective at 0.76.
• A meta-analysis of randomized pre-K–12 trials (Nickow and colleagues, 2024) found 0.29 standard deviations (0.37 in an earlier version), largest with teachers or paraprofessionals as tutors, in earlier grades, at least 3 days a week, during school.
• An analysis of 265 trials (Kraft and colleagues, 2024) found effects only a third to a half as large in studies most like large-scale programs, driven by sharp drops as programs scale up.

Do AI tutors work? What the studies show, and at what age

Some do, in specific designs and mostly with older students.

The Harvard tutor (Kestin and colleagues, Scientific Reports) was built on the same teaching principles as the class; the study says nothing directly about young children. The Nigeria pilot, evaluated by the World Bank, was guided by teachers. In Ghana, schools rather than pupils were assigned, and the authors ask for results to be read "judiciously".

Can an AI tutor harm learning?

Yes, when it does the thinking for the child. In the Türkiye trial (Bastani and colleagues, PNAS, 2025), run at a large high school in 2023–2024, GPT-4 access raised practice grades by 48% with a plain version and 127% with a tutor version. Once access was removed, the plain-version group did 17% worse than students who never had it. The tutor version "essentially eradicated" that harm, but produced no positive effect on exams.

The OECD explains why: "When students depend too heavily on GenAI, metacognitive engagement – the mental processes and effort that turns answers into understanding – drops." The work looks good; the learning does not happen.

What age should children start using an AI tutor?

No official body gives an age, and the evidence thins out the younger the child.

The UK's Best Start in Life guidance tells parents not to let young children use "AI toys, tools or chatbots" until there is more evidence. The only verified AI-learning trial with preschoolers, a story-reading study of 117 children aged 3 to 6, tested comprehension, not tutoring. For younger children's risks, see AI and child development.

How should parents use an AI tutor with a school-age child?

Treat it as a practice partner you supervise, not a homework machine.

1. Choose one that asks rather than tells. Test it: ask for the answer to a homework question. If it simply hands it over, look elsewhere.
2. Try first, then ask. Say: "Show me how you'd start. Then we'll ask it for a hint, not the answer."
3. Stay in the loop. The best results had people involved: teachers in Nigeria, human tutors in Tutor CoPilot. Ask your child to explain what they learned.
4. Do the "without it" check. Every week or so, have your child solve a problem with no AI.
5. Check what it says. Türkiye's MEB advises reviewing AI's suggestions regularly and using varied sources, so a child does not depend on a single AI system.
6. Ask the teacher about the class rules on AI.

What does Türkiye's Ministry of Education say?

MEB is bringing AI into schools with monitoring. Its 2025–2029 AI in Education Policy Document has 4 goals, 15 policies and 40 actions, and says AI's effect on student–teacher interaction will be tracked. Its ethics recommendations ask parents to check that platforms suit the child's age and development, limit use, and explain AI's limits in age-appropriate language.

Where do stories fit in?

MiniMinds is not a tutor, and we make no claim that it teaches school subjects or raises grades. It uses AI to write and illustrate stories for four age bands (3–5, 6–8, 9–11 and 12+), and your child does not chat with an AI. A story for ages 9–11 or a space adventure, read together or in your own voice, is a break from homework. For younger children, our guide to Montessori at home explains why Montessori's founding body keeps screens away from under-sixes.

Sources

1. Bloom BS (1984). The 2 Sigma Problem. Educational Researcher 13(6). DOI 10.3102/0013189X013006004. link
2. von Hippel PT (2024). Two-Sigma Tutoring: Separating Science Fiction from Science Fact. Education Next. link
3. VanLehn K (2011). The Relative Effectiveness of Human Tutoring, Intelligent Tutoring Systems, and Other Tutoring Systems. Educational Psychologist. link
4. Kulik JA, Fletcher JD (2016). Effectiveness of Intelligent Tutoring Systems: A Meta-Analytic Review. Review of Educational Research. link
5. Nickow A, Oreopoulos P, Quan V (2024). The Promise of Tutoring for PreK–12 Learning. American Educational Research Journal. link
6. Kraft MA, Schueler BE, Falken G (2024). What Impacts Should We Expect from Tutoring at Scale? Annenberg Institute, Brown University. link
7. Kestin G et al. (2025). AI tutoring outperforms in-class active learning. Scientific Reports. link
8. Bastani H et al. (2025). Generative AI without guardrails can harm learning. PNAS. link
9. De Simone M et al. (2025). From Chalkboards to Chatbots. World Bank Policy Research Working Paper 11125. link
10. Henkel O et al. (2024). Effective and Scalable Math Support: Evidence on the Impact of an AI-Tutor on Math Achievement in Ghana. arXiv. link
11. Wang R et al. (2024). Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise. arXiv. link
12. OECD (2026). OECD Digital Education Outlook 2026. OECD Publishing. link
13. Xu Y et al. (2022). Dialogue with a conversational agent promotes children's story comprehension via enhancing engagement. Child Development. link
14. UK Government, Best Start in Life (2026). Baby and toddler screen time guidance. DfE and DHSC. link
15. T.C. Millî Eğitim Bakanlığı (2025). Eğitimde Yapay Zekâ Politika Belgesi ve Eylem Planı Yürürlüğe Girdi. MEB. link
16. T.C. Millî Eğitim Bakanlığı (2025). Yapay Zekâ Etiği Tavsiyeleri. MEB. link