Is there any AI tool that genuinely refuses to give answers?

Some hold the line much better than others, and the pressure test in step three is how you find out, because nearly all of them claim to. Treat total refusal as a warning sign rather than a goal — a tool that never helps gets abandoned by the child inside a week, and an abandoned tool teaches nothing. The behaviour worth looking for is the smallest useful step, held under pressure.

My child uses ChatGPT for homework. Is that ruining their learning?

The honest answer is that it depends on a behaviour, not on the brand. The study above found the harm specifically in the arm where the AI would answer; the same model with hints instead did no damage. A general-purpose chatbot answers by default, so the risk is real — but it is a risk about how it is used, which means a house rule (try first, ask for a hint not an answer, explain it back before moving on) changes the exposure more than switching products does.

What age does this start mattering?

As soon as your child can type a question they could have worked out themselves, which for most children is around the start of primary school. Before that the tool is mostly a toy and the question is screen time, not learning. The window where it matters most is the stretch where the habit forms — early enough that nothing is riding on it yet, and before the schoolwork gets heavy enough that the cost of a bad habit becomes visible in a grade.

Can I just tell whether it is working by looking at their schoolwork?

Not reliably, and this is the trap the study illustrates. While the AI was available, the group that was being harmed looked like the group that was doing best — the practice scores went UP. The damage only appeared when the support was removed. So the thing to look at is not performance with the tool, it is whether they can do it without it a few days later.

Does a safety certification tell me anything about this?

No, and conflating the two is the most common mistake in this category. COPPA compliance and content filtering are about what happens to your child's data and what they might be exposed to. Both matter enormously and neither says a word about whether the tool completes your child's thinking for them. A fully compliant product can be a very efficient way to stop a nine-year-old from learning arithmetic.

Parent's guide

AI tools that teach kids to think, not just get answers

The property that separates the two kinds of tool is narrow and testable: does it withhold the answer on the first turn? A tool that answers immediately is a vending machine, however warm its voice; a tool that asks what your child already tried, gives one hint, and makes them finish is doing the thing you are paying for. This is not a matter of taste — in a study of nearly a thousand students, an AI that handed over answers raised practice scores and then LOWERED the same students' later unaided exam scores by about 17%, and a hint-giving version of the same AI erased that damage. Below: the six behaviours that tell the two apart, the ten-minute test you can run tonight on any tool your child already uses, and why the word “Socratic” on a marketing page is not evidence of anything.

The test is withholding, and it is the only one that travels

Ask any AI tool a question in your child's voice — “what is 7 × 14” or “write me a paragraph about volcanoes” — and watch only the first response. If the answer is in it, the tool has taught your child that the gap between not knowing and knowing is one sentence of typing. Everything else on the feature list is downstream of that one behaviour. If the first response instead asks what they have already tried, or offers a smaller step, your child has to do the part where the learning actually happens.

This is worth being blunt about because the category is sold on the opposite axis. Products compete on safety, on filters, on dashboards, on how kid-friendly the character is — all of which matter and none of which touch this. A perfectly safe, COPPA-compliant, parent-approved tool with a cheerful mascot can still be a machine that completes your child's thinking for them, and most of them are, because answering instantly is what the underlying models do by default and withholding has to be deliberately built.

The good news for a parent is that deliberate construction is visible from the outside. You do not need to audit a system prompt. You need one real question and about ten minutes, and the rest of this page is how to spend them.

The side-by-side comparison: eight AI apps for kids · Five checks for whether it is working on your own child

What happened when someone actually measured it

Nearly a thousand high-school maths students were given an AI assistant. One group got an unrestricted version that would answer; one got a version restricted to hints; one got neither. While the AI was available, the unrestricted group performed best — the practice scores went up, which is exactly what a parent looking over a shoulder would see and be reassured by.

Then the AI was taken away and all three groups sat an exam unaided. The unrestricted group scored about 17% LOWER than the students who never had AI at all. The hint-only group scored about the same as the no-AI group: the damage was gone.

Two things in that result deserve care, because it is routinely over-quoted in both directions. The hint version did not produce a gain — it removed a harm, which is a weaker and more honest claim than the marketing version of this study you may have seen. And it was high-school maths in Turkey, not a nine-year-old making a story, so treat it as evidence about a mechanism rather than a prediction about your child. But the mechanism is the one that matters here, and it points at withholding specifically: same model, same subject, same students, and the only variable that moved the unaided score was whether the thing would just tell them.

There is a second finding worth holding alongside it. Independent risk assessments have observed that being good at homework help is precisely what leads young people and their parents to over-trust the same systems on questions they handle badly. Competence in the easy case buys credibility in the hard case. A tool that trains your child to accept fluent output without checking it has not only cost them the practice — it has calibrated them wrongly for the moment the output is confidently wrong.

Six behaviours that tell the two kinds of tool apart

Observable behaviours, not marketing claims — each one you can verify yourself inside a single session, which is why there are no vendor quotes in this table. The third column is what to actually do; the whole exercise takes about ten minutes.

What to look at	Answer machine	Thinking tool	How to check it in under two minutes
The first response	Contains the answer, usually with a tidy explanation after it.	Contains a question back, or one step smaller than the one asked for.	Ask it a question in your child's voice and read ONLY the first reply. Do not reply yet. This single observation sorts most of the category.
What it does under pressure	Gives in. Type “just tell me” twice and the answer arrives.	Holds the line, or narrows the hint without completing it.	Say “just tell me the answer” two or three times. A tool that only withholds until it is asked firmly does not withhold.
Who ends up holding the work	Produces a finished artefact your child collected.	Produces something your child had to assemble, with their own decisions visible in it.	Look at the output and ask whether any choice in it was your child's. If it would be identical for any child, they collected it.
Whether it asks them to explain	Never. The session ends when the answer arrives.	Asks, at least once, how they got there or why it works — before the session ends.	Finish a whole task and see whether anything in the flow asks your child to say it in their own words.
What it does when it is wrong	Is equally confident. Nothing in the design invites your child to doubt it.	Invites checking, and treats finding a mistake as the point rather than a failure.	Feed it something subtly false — a wrong fact in a question — and see whether the design encourages your child to catch it.
What you get to read afterwards	A transcript, an alert, or nothing.	A record of what was worked on, where they got stuck, and how much help it took.	Open the parent view after one real session. The distance between “dashboard” and what is in it closes fast.

The ten-minute test, on whatever your child already uses

1. Pick the question your child would actually ask Not a clever probe — the real one from last Tuesday. A homework question they were stuck on, or “make me a game about cats”. The tool's behaviour on a genuine kid request is the only behaviour that matters, and a test question written by an adult tends to get an adult answer.
2. Read the first response and stop The whole result is usually in there. Answer present: vending machine. Question back or smaller step: worth continuing the test. Write down which, because the next four steps are easier to rationalise away once you are invested in the tool.
3. Push for the answer, twice “Just tell me.” “I don't care, just give it to me.” Plenty of tools have a hint mode that folds the moment a child applies the mildest pressure, and a child will apply far more pressure than you just did. Withholding that only survives politeness is not withholding.
4. Finish one task and look for the explain step Go through to something completed. Did anything, anywhere in the flow, ask your child to say what they made or why it works? If not, there is no quality gate: the tool cannot distinguish a child who understood from a child who pressed go, and neither can you.
5. Open the parent side with that session in it Not the marketing screenshot — the real view, after the real session. You are looking for whether it can answer four questions: what did they work on, where did they get stuck, how much help did it take, and is the thing they got right on Monday still right on Thursday. Most parent views answer none of the four.

Why the word on the marketing page is not the test

“Socratic” has become a product adjective. So have “guided”, “tutor”, “scaffolded” and “never gives the answer”. None of them are regulated and all of them are cheap to type, which is why the test above is entirely behavioural: the claim costs a vendor nothing and the behaviour costs them a real design decision.

There is also a structural reason to distrust the label. Withholding is unpopular in the moment. A child who gets a hint instead of an answer is mildly frustrated, and frustration shows up in engagement metrics within a week — so the commercial pressure on every product in this category runs toward answering, continuously, forever. When you find a tool that genuinely holds the line under pressure, you are looking at a company that chose against its own retention numbers. That is rare enough to be worth checking for rather than assuming.

The flip side, stated plainly because the obvious over-correction is also wrong: a tool that withholds everything is not better, it is just unusable. A seven-year-old who gets nothing but questions quits, and a child who has quit learns nothing at all. What you want is a tool that gives the smallest step that keeps them moving — real choice inside a kept structure, which is the shape that the autonomy-support research has consistently found tracks children's self-regulation. Less help than they want, more than they need to stop.

Where the line sits on homework specifically, subject by subject · The wider roadmap: what to teach, and in what order

Where Luchi sits, with the bias stated

We built Luchi around this exact property, so read this section knowing we have a stake in the test we just handed you. What we do: a child works through a mission rather than an open chat box — they learn one idea, direct AI to build something real with it, hit the point where the AI gets it wrong, and then explain their thinking back. The explain step is not optional decoration at the end; it is how the parent note gets written, in the child's own words rather than ours.

What that means for the six behaviours above: the explain step is built in, the parent side is a record rather than a transcript or an alert, and finding where the AI is wrong is framed as the game instead of as a malfunction.

What we are not: we are not a monitor, we cannot tell you anything about what your child does in a different app or on a school device, and we are not a curriculum that will move a maths grade. If your actual problem is “my child uses ChatGPT on a Chromebook I do not control”, the honest answer is that this page's test is the useful part of it and we are not the purchase. Run the test on what they already use.

Questions parents ask about this

Is there any AI tool that genuinely refuses to give answers?: Some hold the line much better than others, and the pressure test in step three is how you find out, because nearly all of them claim to. Treat total refusal as a warning sign rather than a goal — a tool that never helps gets abandoned by the child inside a week, and an abandoned tool teaches nothing. The behaviour worth looking for is the smallest useful step, held under pressure.
My child uses ChatGPT for homework. Is that ruining their learning?: The honest answer is that it depends on a behaviour, not on the brand. The study above found the harm specifically in the arm where the AI would answer; the same model with hints instead did no damage. A general-purpose chatbot answers by default, so the risk is real — but it is a risk about how it is used, which means a house rule (try first, ask for a hint not an answer, explain it back before moving on) changes the exposure more than switching products does.
What age does this start mattering?: As soon as your child can type a question they could have worked out themselves, which for most children is around the start of primary school. Before that the tool is mostly a toy and the question is screen time, not learning. The window where it matters most is the stretch where the habit forms — early enough that nothing is riding on it yet, and before the schoolwork gets heavy enough that the cost of a bad habit becomes visible in a grade.
Can I just tell whether it is working by looking at their schoolwork?: Not reliably, and this is the trap the study illustrates. While the AI was available, the group that was being harmed looked like the group that was doing best — the practice scores went UP. The damage only appeared when the support was removed. So the thing to look at is not performance with the tool, it is whether they can do it without it a few days later.
Does a safety certification tell me anything about this?: No, and conflating the two is the most common mistake in this category. COPPA compliance and content filtering are about what happens to your child's data and what they might be exposed to. Both matter enormously and neither says a word about whether the tool completes your child's thinking for them. A fully compliant product can be a very efficient way to stop a nine-year-old from learning arithmetic.

Where the evidence on this page comes from

Two sources carry the empirical claims above, and both are cited with the limits that constrain them. Everything else on this page is a behavioural test you can run yourself, which is a deliberate choice: the observable behaviour of a tool in your own session is better evidence for your household than any study average.

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 load-bearing source for the whole page: same model, same students, and the only variable that moved the later unaided exam score was whether the AI would hand over the answer. Read the limits closely before quoting it — the hint-giving arm removed the harm and produced no positive exam effect, and the study is high-school mathematics, so it is evidence about the withholding mechanism rather than a forecast for a nine-year-old. 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. Used for one claim only: that competence at homework help is what buys a system the over-trust it then gets on questions it handles badly. The assessments themselves are written for teens and test general-purpose chatbots, not kid-specific products — only the transfer-of-trust mechanism carries down to the younger children this page is about, not the ratings. 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.
Grolnick & Ryan, 1989 — Parent styles associated with children’s self-regulation and competence in school. Journal of Educational Psychology, 81(2), 143–154. Behind the “smallest step that keeps them moving” point: autonomy support means real choice inside a structure that is kept, not fewer rules. Cited here as the reason a tool that withholds everything is not the better tool — and noting that this study is routinely misread in exactly that direction. What it does not show: Autonomy support means real choice inside a kept structure. The study is not evidence for fewer rules, and it is routinely misread that way.

Run the test on us

Everything on this page is a check you can run on any tool, including ours — and we would rather you ran it than took our word for it. Start a mission with your child, push it for the answer, and see what it does.

Get started

Compare the apps first

Part of the series: Teaching kids about AI

Related guides in this series

How to explain AI to a child — Simple analogies and sample scripts by age (5–12), plus the three things every kid should understand about AI.
AI activities for kids — 10 free, screen-smart activities to explore how AI works together — no coding, no sign-up, mostly off the screen.
Fall break activities for kids — A half-day-by-half-day plan for five unstructured days — ten low-prep blocks, screen and off-screen alternating, ending with your child explaining what they made.
AI tools for kids — A parent’s shortlist of genuinely kid-appropriate AI tools — what to look for, what to avoid, and how to tell a real kid tool from a repackaged chatbot.