Are AI tutors safe for elementary kids?

It depends almost entirely on the product, and the evidence base for this age group is thin enough that anyone giving you a flat yes is overclaiming. Two things are known. An AI that hands over answers can leave a child performing worse on a later unaided test than one who never used it — about 17% worse in the best study available, with high-school students. And general-purpose chatbots earn trust through homework competence that they do not deserve on harder topics. A bounded tool that gives hints, shows you something, and holds a clear consent and data position is a reasonable bet; a general chatbot with a cartoon avatar is not.

Will using AI for homework make my child worse at the subject?

It can, and the mechanism is specific: if the AI supplies answers, your child practises obtaining solutions rather than producing them, and the gap shows up on the next unaided test rather than in the homework. In the clearest experiment on this, the unguarded-AI group scored about 17% below students with no AI at all, while a hint-giving version removed that gap. The practical test at home is not whether the homework looks better — it is whether your child can still do a similar question with the device closed.

How can I tell whether an AI homework helper gives hints or answers?

Test it rather than reading about it. Take a real question from your child’s homework, type it in exactly as they would, and look at the first response. If a complete worked answer comes back without being asked twice, that is the default your child will meet. Check also what happens when they say "just tell me" — a tool that folds under mild pressure is an answer machine with a hint setting.

Is ChatGPT a reasonable homework helper for a 9-year-old?

Not as a default, for two reasons that are different in kind. It answers rather than hints unless carefully steered, which is the arm of the research that performed worse on later tests; and it is a general-purpose assistant, so a child who learns to trust it for fractions has learned to trust it for everything else they might type at ten o’clock at night. Independent assessments of the major assistants found real failures on exactly those harder conversations. With an adult sitting there steering it, it is a different situation — but that is you tutoring, with a tool.

What does COPPA actually require of a kids’ AI product?

As amended and in force since April 2026, the main points are verifiable parental consent before collecting a child’s personal information, a separate consent before disclosing it to third parties for things like targeted advertising, limits on how long it can be kept, and a broader definition of what counts as personal information. Treat it as a floor. A product can satisfy every line of it and still be a poor way for a nine-year-old to learn, which is why the other four questions exist.

Is it cheating if my child uses AI for homework?

That depends on what the AI did, and it is a different question from whether it is safe — worth separating, because the answer to each one does not settle the other. The short version: an AI that explains a method your child then applies is help, and an AI that produces the thing being handed in is not. Where exactly that line sits, subject by subject, and what to do when you think it has been crossed, are the two companion pages in this cluster.

Parent’s guide

What parents should know about AI homework helpers

The honest answer has two halves, and most articles give you only one. There is good experimental evidence that an AI helper without guardrails makes learning worse — measurably, on a later test — while the same model with guardrails does not. And there is almost no evidence about elementary-age children specifically. Here is what is actually known, the five questions to ask any product, and what our own answers are, including where they are weaker than we would like.

The one study every parent should know about

In 2024 a research team ran what is still the most useful experiment here, across nearly a thousand high-school maths students. Three groups: no AI, an unguarded GPT-4 assistant much like the public chatbot, and a version prompted to behave like a tutor — hints rather than answers. Both AI groups did dramatically better on practice problems, the unguarded one by around 48% and the tutor by around 127%.

Then they sat an exam without the AI. The unguarded group scored about 17% WORSE than the students who had never had AI at all. The tutor group’s score was back level with the control — the harm was essentially eliminated, and no benefit appeared in its place.

Sit with how strange that shape is, because it is the single most important thing on this page. The tool that most improved the work most damaged the learning. A parent watching homework get finished faster and more accurately was watching the signal that predicted the worse outcome. Any product telling you that smoother, more successful homework is evidence their AI is working is describing the failure mode.

Two caveats stated plainly, because this page is about not overclaiming. These were high-school students doing maths in Turkey, not nine-year-olds doing a spelling worksheet in Ohio, and the effects do not transfer automatically. And the guardrailed arm showed no positive exam effect — it removed harm. Anyone citing this study as proof that AI tutoring raises attainment is citing the opposite of what it found, including, if we ever do it, us.

The second risk, which is about trust rather than learning

There is a separate failure that has nothing to do with maths. Common Sense Media’s Youth AI Safety Institute, assessing the major general-purpose chatbots for young people, identified a mechanism worth naming: these systems are genuinely competent at homework help and general questions, and that competence leads young people and their parents to extend the same trust to things the systems handle badly — including mental-health and emotional conversations, where the assessments found systematic failures to recognise distress.

The route runs straight through the thing this page is about. Homework help is the on-ramp. A child who has found a chatbot reliable for fractions has no way to know that reliability does not extend to how they feel on a bad night, and nothing in the interface tells them. This is why "is this product good at homework?" is the wrong first question, and why a general-purpose chatbot with a kid-friendly tone is a different kind of object from a tool with a bounded purpose.

The honest limit on that evidence: those assessments are written for teenagers and evaluate general-purpose assistants, not products built for seven-to-twelves. What carries over to a younger child is the mechanism, not the ratings — and arguably it carries over more strongly, since a nine-year-old is less equipped to notice the boundary than a fifteen-year-old.

The same question about ChatGPT specifically, by age

Five questions to ask any AI homework product

1. Does it give hints, or does it give answers — by default? The one with experimental evidence behind it, so ask it first and test it rather than reading the marketing. Type in a real homework question as your child would. If a complete worked answer comes back, you are looking at the arm of that study that scored 17% worse on the exam. "Can be configured to give hints" is not the same as doing it by default, and your child will not be the one changing the setting.
2. Is its purpose bounded, or is it a general chatbot in a friendly coat? Ask what happens when your child types something that is not homework — something sad, something scary, something about a friend. A bounded tool should visibly not be a confidant. This is the trust-transfer risk, and it is the question least likely to be answered on a pricing page. Try it yourself before deciding.
3. What does the parent see, and is it the child’s own words? There are two bad answers. One is nothing — a tool that reports no visibility means you find out at parents’ evening. The other is everything: a full transcript feed is surveillance, and it changes what a child is willing to type. What you want is enough to know whether it is being used as a hint machine or an answer machine.
4. What happens to the data, and who else gets it? The US legal floor is the amended COPPA Rule, in force for all covered operators since April 2026: separate verifiable parental consent before a child’s personal information goes to a third party, limits on how long it is kept, and a broader definition of what counts as personal information. Ask directly whether anything is shared for advertising, and whether the child’s content trains a model. A product that cannot answer that in a sentence has told you the answer.
5. What does it claim, and can the claim be checked? The strongest signal available to you, and it costs nothing. Look for a specific number with a source you can open. "Proven to improve grades" with no study behind it is worse than no claim at all, because it tells you the company is willing to say things it cannot support — about learning now, and about safety later.

What the evidence supports, by claim

The claims you will see on AI-tutoring landing pages, scored against what is actually known for children aged roughly 7–12. We are included, with our own weakest rows left in.

The claim	Status for ages 7–12	What is actually known
Unguarded AI help can hurt learning	Supported, in older students	The clearest result in the area: about 17% worse unaided exam performance in the no-guardrails arm, nearly a thousand high-school students. Not replicated in elementary school.
Hint-style guardrails remove that harm	Supported, in older students	The tutor arm came back level with the control. Note what that is: harm removed, not benefit added.
AI tutoring raises attainment	Not supported	The guardrailed arm showed no positive exam effect. Individual studies of other designs are more encouraging and none of them are about elementary homework help.
AI tutors are safe for elementary kids	Unknown — too broad to answer	"Safe" depends on the product. The independent assessments that exist cover general-purpose chatbots and are written for teens. There is no body of evidence about kid-specific homework tools at this age.
AI chatbots handle a child’s emotional disclosure well	Contradicted	Independent assessment of the major assistants found systematic failures to recognise mental-health distress — which matters here because homework competence is what builds the trust that gets extended to it.
A product is safe because it is COPPA-compliant	A floor, not an answer	Compliance governs consent, retention and disclosure. It says nothing about whether the thing is a good way for a nine-year-old to learn fractions.

The research under this page

Three sources, all with limits that matter for a seven-to-twelve-year-old. This section exists so you can check the two numbers on this page rather than take them from us.

Bastani, Bastani, Sungu, Ge, Kabakcı & Mariman, 2024 — Generative AI without guardrails can harm learning: evidence from high school mathematics. (Nearly 1,000 students; an unguarded GPT-4 assistant raised practice scores but LOWERED later unaided exam scores by about 17%, an effect essentially eliminated by a hint-giving tutor version.) The source of both headline figures: practice performance up, unaided exam performance down about 17% without guardrails, and the harm essentially eliminated with a hint-giving tutor. What it does not show: High-school maths in Turkey, not elementary school, and the "guardrails" arm showed no POSITIVE exam effect either — it removed the harm. Anyone citing this as proof that AI tutoring raises attainment is citing the opposite of what it found.
Common Sense Media, Youth AI Safety Institute risk assessments — Independent risk assessments of major AI chatbots for young people, including the finding that competence at homework help leads teens and parents to over-trust the same systems on things they handle badly. The source of the trust-transfer risk — competence at homework help leading young people and parents to over-trust the same systems on things they handle badly. What it does not show: Assessments are written for teens, and the general-purpose chatbots tested are not the kid-specific products this page is about. The transfer-of-trust mechanism is what carries over to ages 7–12; the ratings do not.
FTC, amended COPPA Rule — The Children’s Online Privacy Protection Rule as amended (published 22 April 2025; full compliance required from 22 April 2026), adding separate verifiable parental consent for third-party disclosure, data-retention limits, and a broader definition of personal information. The legal floor behind question four: separate verifiable parental consent for third-party disclosure, retention limits, and a broader definition of personal information. What it does not show: A legal floor for US operators, not a quality bar. A product can be fully COPPA-compliant and still be a bad way for a nine-year-old to do maths.

Our own answers to the five questions

We sell one of these, so the five questions apply to us and it would be a strange page that ducked them. Taking them in order, and no more strongly than is true.

Hints, not answers, by default — that is the design and there is no setting a child can flip to get a worked solution instead. On purpose: the study above is the reason the product works this way rather than a fact we found afterwards. Second, the purpose is bounded. Luchi is a creation and homework coach, not a confidant; talk and games close at bedtime and at the daily limit you set. Third, you see your child’s own words rather than a transcript feed — the parent note is built from what was actually done, which is enough to tell a hint habit from an answer habit without reading over their shoulder. Fourth, consent is parent-first by design and no child uses Luchi without a parent account and consent on file; our privacy policy is the document to hold us to, not this paragraph.

The fifth is the one where we have to be least comfortable, and leaving it out would fail our own test. We have no study of Luchi. Nothing on this page is evidence that our product improves learning, because that evidence does not exist — what exists is a design choice aligned with the one good experiment in the area, which is a reason to prefer hints over answers and not a result. If you ever see us claim a learning outcome without a study you can open, hold us to this paragraph.

And the genuinely honest version of the advice: for a seven-to-twelve-year-old doing ordinary homework, the question "which AI helper?" is less important than whether the homework is theirs and whether anyone is giving them answers. A child with a clear plan and no AI is in better shape than a child with an excellent AI finishing their worksheet for them.

Where the line sits between help and doing it for them · Our privacy policy — the document to hold us to

Questions parents ask

Are AI tutors safe for elementary kids?: It depends almost entirely on the product, and the evidence base for this age group is thin enough that anyone giving you a flat yes is overclaiming. Two things are known. An AI that hands over answers can leave a child performing worse on a later unaided test than one who never used it — about 17% worse in the best study available, with high-school students. And general-purpose chatbots earn trust through homework competence that they do not deserve on harder topics. A bounded tool that gives hints, shows you something, and holds a clear consent and data position is a reasonable bet; a general chatbot with a cartoon avatar is not.
Will using AI for homework make my child worse at the subject?: It can, and the mechanism is specific: if the AI supplies answers, your child practises obtaining solutions rather than producing them, and the gap shows up on the next unaided test rather than in the homework. In the clearest experiment on this, the unguarded-AI group scored about 17% below students with no AI at all, while a hint-giving version removed that gap. The practical test at home is not whether the homework looks better — it is whether your child can still do a similar question with the device closed.
How can I tell whether an AI homework helper gives hints or answers?: Test it rather than reading about it. Take a real question from your child’s homework, type it in exactly as they would, and look at the first response. If a complete worked answer comes back without being asked twice, that is the default your child will meet. Check also what happens when they say "just tell me" — a tool that folds under mild pressure is an answer machine with a hint setting.
Is ChatGPT a reasonable homework helper for a 9-year-old?: Not as a default, for two reasons that are different in kind. It answers rather than hints unless carefully steered, which is the arm of the research that performed worse on later tests; and it is a general-purpose assistant, so a child who learns to trust it for fractions has learned to trust it for everything else they might type at ten o’clock at night. Independent assessments of the major assistants found real failures on exactly those harder conversations. With an adult sitting there steering it, it is a different situation — but that is you tutoring, with a tool.
What does COPPA actually require of a kids’ AI product?: As amended and in force since April 2026, the main points are verifiable parental consent before collecting a child’s personal information, a separate consent before disclosing it to third parties for things like targeted advertising, limits on how long it can be kept, and a broader definition of what counts as personal information. Treat it as a floor. A product can satisfy every line of it and still be a poor way for a nine-year-old to learn, which is why the other four questions exist.
Is it cheating if my child uses AI for homework?: That depends on what the AI did, and it is a different question from whether it is safe — worth separating, because the answer to each one does not settle the other. The short version: an AI that explains a method your child then applies is help, and an AI that produces the thing being handed in is not. Where exactly that line sits, subject by subject, and what to do when you think it has been crossed, are the two companion pages in this cluster.

Hints, not answers — by design

Luchi helps with homework by asking the next question rather than handing over the solution, and there is no setting a child can flip to get the answer instead. You see what got done in their own words. We make no claim that this raises grades, because no study of Luchi exists — the design follows the only good experiment in the area.

Get on the list

How Luchi teaches

Part of the series: How to end the daily homework battle

Related guides in this series

Teaching kids to start homework without reminders — Independence is a prompt you remove on purpose, not a trait you wait for. A five-rung ladder from "you say it daily" to "you only ask afterwards", with how long each rung takes.
Seven practical homework focus strategies — Seven things that help once they have started, in the order worth trying them — plus an honest table of which rest on solid evidence and which do not.
AI homework help for kids — Where the line sits between help and doing it for them, subject by subject — plus a five-step routine for one hard question and five questions to ask any homework product.
Is my kid cheating with AI? — Why AI detectors are not reliable enough to accuse a child on, the ninety-second check that is, and what to do about each of the four things you are likely to find.
The after-school hour, without the nagging — The fight is about the forty minutes before homework. A five-block plan, the four traps that turn a routine back into nagging, and the research that makes it work.