AI tutors for kids: what the research says works

Some AI tutors raise test scores. Others make kids worse at the test. Three design choices decide which one you get.

By Lionel Mora · October 14, 2026 · 5 min read

The short answer

AI tutors work under three conditions. They give hints, not answers: in a 2025 PNAS study, plain GPT-4 made students 17% worse on the exam, and a hint-based tutor avoided the drop. Kids use them regularly: in a two-year trial, students used Khanmigo on only about a third of their Khan Academy days, and gains matched Khan Academy without AI. An adult stays in the loop: a World Bank pilot in Nigeria, with teachers supervising, produced large gains in six weeks.

Human tutoring raises scores by about a third of a standard deviation on average. That's the realistic bar for AI. Bloom's famous 2 sigma is not.
Answer machines hurt learning. Hint machines help.
Explaining is the easy part for a tutor. Getting a 9-year-old to come back tomorrow is the hard part.
For ages 6 to 12, pick a tool built for that age, with a parent account and a clear curriculum.

The pitch for AI tutoring is simple: a patient, expert tutor for every child, any time, for the price of an app. I understand why parents want that. I'd want it too. Does it work? It depends almost entirely on how the tutor is designed.

The benchmark: what human tutoring achieves

In 1984 the psychologist Benjamin Bloom reported that students tutored one to one scored about two standard deviations above students in a normal class. Every edtech pitch since has quoted that "2 sigma" figure. It doesn't hold up well. The original studies used narrow tests and highly trained tutors, and later meta-analyses find effects closer to a third of a standard deviation¹.

A third is still a lot. A review of 96 randomized trials of tutoring found an average gain of 0.37 standard deviations, with bigger effects when tutors were trained, in earlier grades and during the school day². So forget "two grades ahead overnight". The fair test for an AI tutor is whether it can match a good human one.

What the AI tutoring studies show

Study	Who	Result
Bastani et al., PNAS 2025³	About 1,000 high school math students, Turkey	Plain GPT-4: practice scores up 48%, exam scores down 17%. Hint-based tutor: practice up 127%, exam drop largely avoided.
World Bank pilot, Nigeria 2024⁴	Secondary students, after school, with teachers supervising	About 0.3 standard deviations in six weeks, which the researchers put at close to two years of typical learning. Girls gained more.
Kestin and Miller, Scientific Reports 2025⁵	194 Harvard physics students	The AI tutor roughly doubled learning gains compared with an active-learning class, in less time.
Khanmigo trial, NBER 2026⁶	18 Tennessee middle schools, two years	Gains similar to Khan Academy without AI. Students used Khanmigo on about a third of their Khan Academy days.

Look at who's missing: 6 to 12-year-olds. Almost all the good evidence comes from teenagers and university students. AI tutoring could still work for younger kids, but the design has to do more of the work.

The three things that decide whether it works

1. Hints, not answers

The PNAS study is the clearest result in the field. Same model, same students, same math. The version that handed out answers made kids feel better during practice and do worse on the test. The version teachers designed to give hints didn't have that problem³. Decades of learning research point the same way. Trying a problem before being shown how beats being shown first⁸, and pulling an answer out of your own memory makes it stick far better than re-reading it⁹.

In practice: when your child is stuck, a good tutor asks a smaller question. A bad one finishes the problem.

2. Kids have to use it

Nobody puts this finding on the box. In the two-year Khanmigo trial, the AI was right there, but students used it on only about a third of the days they spent on Khan Academy, and the gains were no bigger than Khan Academy alone⁶. A tutor that explains beautifully is useless if a 9-year-old won't open it.

Adults use tutors because they have a goal. Kids use things because they want to. So an AI tutor for children has to solve the motivation problem first. If it doesn't, it never gets a shot at the math problem.

3. An adult in the loop

The biggest real-world gains so far, in Nigeria, came with teachers in the room guiding how students used the AI⁴. The tutor did the patient, one-on-one explaining. The adult kept it on track. At home, you don't need to sit next to your child for every session. You need to know what they worked on, and ask them about it.

How to choose an AI tutor for your child

Is it built for their age? General chatbots start at 13. A 7-year-old needs something made for 7-year-olds.
Give it a wrong answer. Type or say something wrong on purpose. Does it ask how you got there, or just correct you?
Ask it for the answer straight out. Does it hand it over?
Does it follow a curriculum? Practice at your child's level, on what their grade covers, beats random practice.
Will your kid pick it on their own? Try it for a week without nagging. How often they open it tells you more than any review.
Can you see what happened? A short report of what they worked on is how you stay the adult in the loop.
Is it safe? Parent account, filtered content, no ads, no data sold or used for training without consent. Our AI safety checklist goes deeper.

How Marble approaches it

We built Marble around the engagement problem, because I think it's the real one. Marble is an interactive learning world for kids aged 6 to 12: an island with quests, crafting, drawing and characters. Today it's about science, making things, creativity and social skills, not math drills. Kids come back because it's a game they want to play. Inside it, the characters ask questions and give hints instead of handing over answers, aiming just past what each child has already mastered. After every session, a parent gets a short note on what their child did and what to ask them about.

That's a different bet from a homework chatbot. My view: a tutor a child wants to visit beats a better tutor they avoid.

Questions parents ask

Do AI tutors work?

The best evidence says they can. In a World Bank pilot in Nigeria, GPT-4 tutoring with teachers supervising raised scores by about 0.3 standard deviations in six weeks. In a Harvard physics study, an AI tutor roughly doubled learning gains compared with an active-learning class. But an AI that just hands out answers can make students worse: a 2025 PNAS study found a 17% drop on the exam afterwards.

What is the best AI tutor for kids?

Pick one built for your child's age that gives hints instead of answers, follows a real curriculum, is set up through a parent account, and that your child opens without a fight. For ages 13 and up, Khanmigo from Khan Academy costs $4 a month. For ages 6 to 12, Marble is an interactive learning world about science, making things, creativity and social skills, where the characters answer kids with questions and hints rather than answers.

Is an AI tutor as good as a human tutor?

Not yet, in general. A review of 96 randomized trials found human tutoring raises achievement by about 0.37 standard deviations, most of all with trained tutors during the school day. AI tutors in good studies have come close to that, but mostly with older students and a teacher supervising.

Should my child use ChatGPT as a tutor?

Not if they're under 13, which is ChatGPT's minimum age. Even for teens, plain ChatGPT tends to give the full answer. If your teen uses it, turn on Study Mode and have them explain every answer back to you in their own words.

How much does an AI tutor cost?

It varies a lot. Khanmigo costs $4 a month or $44 a year for families. Many subscription learning apps with AI features cost $10 to $20 a month. A human tutor in the US usually costs far more per hour, which is a big reason AI tutoring is attractive.