Every AI app for children says it is safe. The word does a great deal of work and almost none of it is regulated — there is no test you must pass before writing “safe for kids” on a product page. So the useful question isn't whether an app claims to be safe. It's what the app actually does, and what the people who built it will tell you when you look.
1. Which AI model is behind it, and will they name it?
Almost every kids' AI app is a wrapper around a general-purpose model — GPT, Gemini, Claude, Llama. There is nothing wrong with that; building a model from scratch for a children's app would be worse, not better. What matters is whether the company says which one.
An app that won't tell you what is generating the answers has told you something already.
2. What happens to your child's questions after they're answered?
The model doesn't run on the phone, so the question goes somewhere. That much is true of every AI app, including ours. The things worth knowing are narrower:
“We don't sell your data” answers none of these. It is the easiest sentence in the industry to write.
- Is the question stored after the answer comes back, or discarded?
- If stored, for how long, and attached to what — a device, an account, a name?
- Is it used to train the model?
- Does the AI provider get it under their own terms, or under a business agreement that forbids training on it?
3. Is filtering done before or after the model answers?
Filtering only the child's question means the model can still produce something you wouldn't want, because nothing checks the answer. Filtering only the answer means unsafe input reached the model in the first place. Doing both is meaningfully harder and meaningfully better.
Ask which one an app does. Most marketing pages say “filtered” and stop there.
4. Does it show you what your child asked — and do you want that?
This one cuts both ways, and most articles like this pretend it doesn't.
An app that reports every question to a parent dashboard gives you real oversight. It can also be the reason a ten-year-old stops asking the questions they most need answered, because they know you'll see. Children ask AI things they won't ask you — that is precisely the value and precisely the risk.
There isn't a correct answer here. But decide what you want before you buy, rather than discovering the app's opinion afterwards.
5. Does it need an account, and what gets attached to it?
An account means a database somewhere with your child in it: a name, an email, probably an age, and a history. That database can leak, be sold with the company, or be subpoenaed.
Apps that work without an account can't lose what they never had. If an app does require one, ask what it needs it for — sometimes the answer is genuinely syncing between devices, and sometimes it's a marketing list.
6. If it's free, work out who is paying
Answering a question with a large AI model costs the developer real money every single time. A kids' AI app with no subscription, no limits and no ads is being paid for by something you haven't been told about yet.
Ads in children's apps mean tracking. Free with a daily limit is usually honest. Free with no limit at all deserves a hard look.
7. What does it do when it doesn't know?
Language models invent things fluently and without warning. For an adult that's an annoyance; for a nine-year-old with no way to check, it's misinformation delivered by something that sounds like a teacher.
Test it before you hand it over. Ask about something obscure and local — a small museum near you, a minor historical figure. See whether it says it doesn't know, or produces a confident paragraph of fiction.
No app can promise a model will never do this. Anyone promising that is overselling.