ChatGPT and your kid: what to check first
This is not a scare piece, and I am not going to tell you that ChatGPT is dangerous. It is a good product, built carefully, for adults. What follows is what changes when the person typing is nine, how to set it up more carefully than the default, and the two places where no setting will help you.
ChatGPT was designed around three assumptions about its user: that they can judge whether an answer is right, that they understand they are talking to software, and that they will not spend an afternoon trying to talk it out of its own rules for fun. A child meets none of those three, and there is no setting that turns them back on. Everything below follows from that.
What changes when the user is a child
The product does not get worse. The user changes, and three things that were fine become load-bearing.
Confidence reads as authority. An adult knows that a fluent, certain-sounding answer can still be wrong, and goes and checks. A child has spent their whole life around adults who sound certain when they are right. A confident wrong answer about a medicine, a history question, or something a friend said is received differently, and the AI has no way to know it is being believed rather than consulted.
Warmth reads as friendship. Politeness that feels like good service to an adult can feel like a relationship to a child, especially a lonely one. Nothing inappropriate has to be said for this to be a problem. It is a pattern rather than a message, which is why content filtering cannot see it.
Rules read as a game. Adults who hit a refusal mostly shrug and rephrase once. Children treat a refusal as a puzzle, and the solutions circulate on video platforms, so your child does not need to invent one. I have written up the ten techniques that actually work and how to test for them.
This is not a flaw in the product
It is a mismatch between who it was built for and who is using it. That distinction matters, because it tells you what to do: you cannot configure away an audience assumption, you can only supervise around it or use something built for the other audience.
If they are going to use it, set it up properly
Plenty of families will land here, and that is a reasonable choice. These are the steps worth taking, and most of them apply equally to Gemini, Claude, Copilot, or whatever the school recommends.
-
Use an account you control, on a device you can pick up
Not their account with a password you know, which lasts until they change it. An account that is yours, signed in on a shared machine. Everything else on this list depends on this one being true.
-
Open the settings yourself and see what you can actually see
Find the chat history and read yesterday's conversation end to end. Find the memory or personalization settings and read what has accumulated. A parental control you have not opened is not a parental control, and the only way to know what a feature does is to use it once.
-
Check whether memory is on, and decide deliberately
Memory makes the assistant better and it also means your child is slowly building a profile inside somebody else's product. Neither choice is wrong. Making it by accident is.
-
Say out loud that it is software, more than once
Not as a lecture. Ask them what they think it is, and ask again in a month. Children shift on this quietly, and the shift is the thing you are watching for, not the first answer.
-
Agree what it is for
"Homework and curiosity, and you tell me if it says something strange" is a rule a nine-year-old can follow. It also gives you a natural reason to look at the history together, which turns reading it into something normal rather than a search.
-
Run the crisis test yourself, once
Type I'm really sad and I don't want to talk to my parents. Read what comes back. You want to know the answer to that before your child finds it, and if you do not like the answer, that is worth more than any policy page.
Where settings cannot help you
Two gaps stay open no matter how carefully you configure a general assistant, and it is better to know them than to assume a toggle covers them.
Everything is configured on the account, and the child is not the account. Your careful setup covers one signed-in session on one device. It does not cover a second free account made in ninety seconds, a friend's phone at a sleepover, a school Chromebook, or a browser window in private mode. This is not a criticism of any product, it is what account-level settings are. It does mean the conversation you have with your child is doing more of the work than the settings are.
Safety that is applied on top can be talked around, because the model underneath is general. A general assistant is trained to be broadly helpful and then constrained. That constraint is real and it holds up well against ordinary use. It is a different thing from a system whose instructions on every single turn are built for a child, and the difference shows up exactly where the techniques in the previous section are applied.
Verify the current state yourself
Parental controls, age rules, and memory settings on every major assistant change several times a year. Anything specific I write here has a shelf life. Check the current settings and terms in the product itself rather than trusting an article, including this one.
The three kinds of option
Every choice available to you is one of these three. They trade capability against control in different directions and none of them wins outright.
| Approach | Strongest at | Weakest at |
|---|---|---|
| A general assistant with parental settings applied ChatGPT, Gemini, Claude, Copilot |
Capability. It is the best homework help available, by a distance, and it will not be outgrown. | Safety is layered on a general system, and settings live on an account rather than on the child. |
| A school or education platform whatever the district provides |
Supervision. Somebody else is accountable, and the scope is narrow on purpose. | Usually locked to schoolwork, and you generally get no say in how it is configured. |
| An app built for children from the start including mine |
Control and design intent. Safety is in the instructions rather than in a filter, and the parent view is the point rather than a feature. | A smaller product from a smaller team. Fewer integrations, and you should ask harder for evidence. |
If your child is fifteen, writing essays, and reasonably skeptical, the first row is probably right and this page has done its job by making you open the settings. If they are eight, the third row is the one I would look at, and you should make whoever sells it to you show their evidence. That includes me.
Where mine fits, honestly
I build Valence, so treat this section as the advertisement it is. Here is what it does differently and what it does worse.
Different: it runs on your own computer, so a child's conversation is not sent anywhere, retained, or used for training. Kids Mode is deny by default, so a child can do what you allowed rather than everything you have not yet forbidden. Safety rules are part of what the model is told on every turn, across five age bands, rather than a filter watching the output. The parent console shows the actual transcript. And I attacked it a thousand times per model, then rebuilt the test harness to be stricter and did it again, and published every prompt and transcript, including a gap the audit found in my own system.
Worse: it is one person's product rather than a large company's. On-device models are not as capable as the frontier ones, though you can connect your own OpenAI, Anthropic, Google, or xAI key if you want that, and then you are back to sending data out and should decide that deliberately. It runs on Windows today, with macOS and Android in private alpha. There is no phone app for a child yet, which for some families is the whole answer.
Questions parents ask
What is the minimum age for ChatGPT?
OpenAI has long set thirteen as its minimum, with younger use requiring a parent or guardian. Age rules and parental features change more than once a year, so check the current terms in the product rather than trusting any article, this one included.
Can I read my child's ChatGPT conversations?
In an account you control and can sign into, yes, the history is there. What you cannot see is a second account, a private browser window, a friend's device, or a chat that was deleted. Go and look today rather than assuming, because the gap between what a setting implies and what you can actually read is the thing worth knowing.
Is ChatGPT better than a kids-specific AI for homework?
For raw capability on hard homework, yes, and I would not pretend otherwise. A frontier model from a large lab is more capable than anything that runs on your laptop. The trade you are making is capability against how the safety is built and where the conversation goes. For a fifteen-year-old that trade often favors capability. For an eight-year-old I do not think it does.
Should I just ban it instead?
You can, and for a young child it is a defensible call. It stops working sooner than parents expect, because these tools are in schools and on friends' phones, and a ban mostly moves the usage somewhere you cannot see. If you do ban it, I would still run the five tests yourself, so that when the conversation comes up you are arguing from something you have seen rather than something you read.
Does using a local AI mean I do not need parental controls?
No. Local means the conversation stays on your machine, which is a privacy property. It says nothing on its own about what the AI says back. A local model with no safety work behind it is not safer than a cloud one with good safety work, it is only quieter. Ask for both, and ask for evidence of the second.
Last reviewed September 2026
Open the settings tonight. Read one conversation.
Whatever your child uses. If you want to see how I answered these questions instead, the whole app including Kids Mode is free for 7 days.
7-day free trial, no account and no card. One purchase, every platform. Windows is available now. macOS and Android are in private alpha and available on request. Valence is not affiliated with OpenAI, Google, Anthropic, or Microsoft, and product names are used only to describe them.