Some AI tutors help. Most of what you can buy has never been tested. The research that does exist is on tutors a teacher built to withhold the answer, and it shows the same underlying tool can add a lot or take learning away depending on that one design choice. So before you pay for anything, check whether the tutor in front of you can refuse to answer.
Do AI tutors actually work?
Two trials have tested this properly. They produced opposite headlines, and read together they say the same thing.
Both were randomized trials: students were put into the different conditions by chance, not by choice. That matters here, because the children most likely to be handed a tutoring app are the children already behind, so a study that let families opt in would be measuring the families, not the tutor.
The first ran at Harvard and was published in Scientific Reports on June 3, 2025, by Greg Kestin and colleagues in the physics department. Each of 194 undergraduates took one lesson with an AI tutor and one in the usual class. After the AI lesson the median score was 4.5 out of 10, against 3.5 after the class lesson, from a starting point of 2.75, and they got there faster: a median of 49 minutes against roughly 60. They also reported being more engaged. Hold it lightly, though. That is Harvard undergraduates on two physics topics, and nobody went back later to see what they still remembered.
The second was published in PNAS by Hamsa Bastani and colleagues, and it ran in fall 2023 with nearly a thousand students in grades 9 to 11 at a high school in Turkey, across four 90-minute math sessions. There were three groups: ordinary study materials, a plain ChatGPT-style assistant the authors called GPT Base, and a version called GPT Tutor that the students' own teachers had configured to give hints instead of answers, to work one step at a time, and to carry the teachers' worked solutions and the mistakes students usually make.
While the AI was in front of them, both groups did better on the practice problems, and the tutor version did far better: 48 percent above the control group for GPT Base, 127 percent above it for GPT Tutor. Then the students sat an exam with no AI at all. The GPT Base group scored 17 percent worse than the students who had never been given it. The GPT Tutor group came out level with them. The authors' own conclusion is that unrestricted access can harm educational outcomes, and they describe the students on the plain version as using it as a crutch.
That is the evidence. Here is what I make of it.
The two trials do not contradict each other. In both, the thing that worked was not the AI. It was a teacher writing rules that stopped the AI handing over the answer. At Harvard the instructors wrote prompts for each question and gave the model step-by-step answer keys, which is also the only reason it stayed accurate. In Turkey the teachers did the same with hints and common errors. Neither of those tutors was a product a family could buy.
And notice what the Turkish result actually is, because no advertisement is going to tell you. The well-built tutor did not beat ordinary studying on the later exam. It drew level. Its achievement was to stop the AI doing damage. That is worth having, and it is much less than the word tutor promises.
The one test to run before you pay
The Withholding Test is the one check to run before you pay for an AI tutor: ask your child to get the final answer out of it in one move, in front of you. If it hands the answer over, you have not bought a tutor. You have bought a faster way to finish.
Use a question your child has already done, so you can also see whether the answer is right. Most products fail this in under a minute. Some hold out for one round of hints and then fold when the child says they are stuck. A few genuinely will not do it, and those are the only ones the research is about.
An AI tutor is not a better answer machine. It is an answer machine with the answers taken away. If yours will not withhold, it is not tutoring your child. It is finishing their homework politely.
Is an AI tutor as good as a human tutor?
Nothing sold to families has the evidence base that human tutoring has. The Education Endowment Foundation, which rates the evidence behind school interventions in England, puts one to one tuition at five months of additional progress, on moderate evidence strength and at moderate cost.
The conditions the EEF attaches to that number matter more than the number. Short, regular sessions of about 30 minutes, three to five times a week, over a set period of up to ten weeks. Strongest results when the tutoring is delivered by experienced and specifically trained teachers. And tuition that is additional to, but explicitly linked with, normal teaching, with progress monitored.
Read that list again and most of it has nothing to do with who does the tutoring. Thirty minutes, three times a week, for ten weeks, on a topic somebody chose, checked against what is happening in class: that is the intervention, and the tutor is only the part that delivers it. Which is the most practical thing I can tell you here. You cannot buy an AI tutor with the EEF's evidence behind it. You can copy the EEF's conditions for nothing.
What about the products built as tutors?
Some are built to withhold, and those are the ones worth your time. Khan Academy's Khanmigo is the clearest example. Its page for parents says it "doesn't just give answers" and instead "gently guides them to find the answers themselves"; it lists at $4 a month or $44 a year on a parent subscription, in US dollars on the US site; the account holder has to be 18 or over, and a parent sets up and manages a child's account (khanmigo.ai, read September 13, 2026). It is tied to Khan Academy's own courses, which is a real limit and also the reason it knows which question your child is on.
Those are the company's facts about its own product, its claim about how it tutors included. They are not evidence that it works, and I have not tested it. Run the Withholding Test on it yourself, the same as on anything else.
As a rule of thumb, a tutor built around one subject holds the line better than a general chatbot, because somebody had to write the steps. That is most visible in math, or maths as it is in the UK, where the steps are the thing being learned.
What I See in Schools
When a school puts money into tutoring, the software is rarely the part that decides whether it works. The pattern I see most often is that the schools getting something out of it are the ones where somebody sat down and decided which handful of topics a particular group of children would cover, in what order, and who would check. The schools getting nothing are the ones where the tool was bought and the children were pointed at it.
At home that decision almost never gets made. The app gets bought, the login goes to the child, and the hope is that the software will work out what is missing. It will not. It has no idea what your child does not know.
In the advisory work I do with education departments, nobody asks which tutoring product is best. They ask what it is meant to be replacing. It is a better question, and it works just as well at a kitchen table.
What to do this week, by age
Under 8. Not a chatbot. OpenAI's own help pages say ChatGPT is not meant for children under 13, and that in education settings with that age group an adult has to do the interacting (OpenAI Help Center, read September 13, 2026). For early reading and number work, the thing with evidence behind it is an adult sitting alongside the child. If you want AI here, use it yourself: have it make you a set of practice questions at the right level, then do them together.
8 to 12. On your account, with you in the room, on one topic their teacher has named. Run the Withholding Test together and let them watch it fail, because that is the lesson. My guide to AI for 9 to 12 year olds goes further into which account a child this age should be on.
You could say: "Ask it to help you with this one, but tell it not to give you the answer. If it gives you the answer anyway, come and tell me, because that means it is not being a good teacher."
13 to 15. ChatGPT's stated minimum age is 13, and OpenAI asks for a parent's consent for users aged 13 to 18. Age is self-declared at sign-up, so the age limit is a promise, not a lock. This is also the band where "help with my homework" quietly becomes "do my homework", and the two look identical in a browser history. Agree what the tutor is for before you agree to pay for it, and set up the parental controls while you are there: how to do that, and what they cannot do.
You could say: "I do not mind you using it for math. I mind you using it instead of math. Show me one where it got you unstuck and you finished it yourself."
16 to 18. Hand over the decision and keep one check, the one I call the Clock Test: has the work got faster without getting harder? If a task that used to take forty minutes now takes eight and the grades went up, the effort moved to the machine, and the learning went with it. It is not a verdict on its own, because fast is sometimes just fluent. It is the question behind what the research says about AI and children's thinking, and it is where the international picture sits too. The OECD's PISA 2025 results, published on September 8, 2026, found that students who do not use AI for specific schoolwork tasks outperform those who do. They also found that students "tend to perform slightly better when frequent AI use is paired with opportunities to develop AI literacy skills".
The prompt that does most of the work
If you are using a general chatbot rather than a dedicated product, you can build the guardrails the trials tested. Paste this at the start and fill in the blanks:
You are tutoring a student of [age] in [subject]. Never give the final answer. Ask one question at a time. Offer a hint, then wait for their attempt. If they get it wrong, tell them which step went wrong and ask them to try that step again. Keep every reply to two or three sentences.
That is not a trick. It is a plain version of what the teachers in both trials did, and it costs nothing. It also wears off: a determined 15 year old can open a new chat. The prompt sets a tutor up. It does not police one.
What an AI tutor cannot do
It cannot tell you why your child is stuck. Working out that a 12 year old's math problem is really a reading problem is a judgment a teacher makes, and it is the reason to ask the school before you buy anything.
It will still be confidently wrong. The Harvard tutor stayed accurate because the instructors fed it the worked solutions. Yours has not been given those, so the answers in the subjects you cannot check yourself are exactly the ones to treat as a claim rather than a fact.
It is not an assessment. If you suspect a specific difficulty with reading, writing or number, the route is the school's special education process, called SEND in the UK, and a professional assessment. No app substitutes for that, and starting with the app usually costs you a term.
And it does not overrule the school. If your child's school has a policy on AI in coursework, or an exam board has rules about what help is allowed on assessed work, those come first, whatever any product says. Where help ends and doing begins is the test teachers actually apply.
Before you pay, ask the school one question
Which two or three topics would actually move my child on, and in what order?
A good answer names topics. A weak answer names a subject. If you get topics, you have the ten-week plan the EEF's evidence is built on, and any decent tutor, human or otherwise, can deliver it. If you get a subject, you have just learned that nobody has diagnosed anything yet, and knowing that is worth more than the first month's subscription.
One more thing worth doing whatever you decide: do your own checking out loud in front of your child. A child who has watched a parent catch an AI being wrong has learned something no tutor can sell them.
If your child's school has not yet spoken to parents about any of this, it is the work I do with schools: training teachers on AI in the classroom, setting the strategy, and making sure the policy means something where your child is sitting. It is worth forwarding. And if you want this sort of thing without the panic, I write a Sunday newsletter that keeps teachers and parents up to date. Parents are very welcome.
Dan Fitzpatrick trains teachers and advises education departments on what AI means for children's learning. More about Dan.
