Open any children's app in the store, scroll past the screenshots, and read the data-safety section. It takes about forty seconds, and it usually tells you more than the entire feature list above it: whether the app collects a name, whether it asks for location, whether anything is shared with other companies, whether the developer says data can be deleted on request. Most parents skip that panel because it reads like paperwork. It is the one part of the listing the developer had to be careful about.
Safety in an AI product for children is structural. It is not a friendly tone or a cartoon guide with a soft voice. It is a set of decisions about what the software is capable of doing at all, and most of them are visible from outside once you know where to look.
Free-form chat is the first thing to rule out
A general-purpose assistant follows a conversation wherever it goes. That is the entire point of it. Hand the same interface to a nine-year-old at eight in the evening and the conversation goes somewhere nobody planned: a question about a pet that died last year, a rumour from the playground, the spelling of a word they heard in a song. The model may handle it gracefully. It may also produce something confidently wrong, or emotionally heavy, with no adult in the room.
This does not mean AI has no place in a child's learning. It means the AI belongs on the generation side rather than the conversation side. A system can use a model to write the next set of reading questions at the right difficulty, to decide which multiplication facts a child should see again on Thursday, or to score a child reading a passage aloud, without ever putting an open text box in front of them. The child taps, chooses, spells, reads and answers. Kid Genius World is built on that pattern, and so are most products in this category that take the question seriously.
If an app does have a text box, spend five minutes with it yourself before your child does. Type what children actually type: i hate school, you are stupid, where do babies come from, my grandma died. A carefully built product will redirect calmly to the lesson and, at best, flag to the parent that something off-topic came up. A carelessly built one will improvise. You will know which one you have within three messages.
Cameras, photos and microphones
Image upload is the feature that most often turns a tidy app into a privacy problem. Children do not upload the homework page. They upload their own face, a sibling, the inside of a bedroom, and fairly often a worksheet with their full name and the school's name printed across the top. Once that image leaves the tablet, the question is no longer about the app but about wherever it is stored and whoever it is passed to.
Microphones sit differently, because reading practice genuinely needs one. A pronunciation check has to hear the child. The fair question is what happens to the audio afterwards: processed on the device or sent to a server, held for a few seconds or held indefinitely, used to improve the product or used only to score that one attempt. "The recording is processed to score the reading and is not retained" is a reasonable answer. "Recordings may be used to improve our services" means retained, and possibly reviewed by people.
Other users are the biggest single variable
An app with no other children inside it has an entire category of risk removed, not reduced, removed. There is no contact surface, no bullying surface, no moderation queue that might be thinly staffed on a Sunday night, no display name that quietly leaks a real name and a birth year.
The feature to look for is not only chat. Global leaderboards show your child a stranger's chosen name. Shared galleries of drawings or stories show your child text another person wrote. Even fixed-phrase chat, where children can only send sentences from an approved menu, has been used for exclusion in plenty of games. One question covers all of it: is there any screen where my child sees content another user created? If the answer is yes, the follow-up is who reviews that content and how quickly.
What is actually being collected
Almost every app collects something, and some of it is genuinely necessary. The useful skill is telling the categories apart.
| Category | Why an app might legitimately need it | What should make you pause |
|---|---|---|
| Parent account details | Login, billing, password recovery | Marketing consent bundled into signup with no way to decline |
| Child profile | A first name or nickname and an age band, so content lands at the right level | Full name, date of birth, school, photo or home address |
| Learning records | Which items were answered, right or wrong, and when. This is what makes practice adapt | Little, as long as it stays tied to the account and can be deleted |
| Device and diagnostics | Crash reports, device model, app version | Precise location, contact list, a list of other installed apps |
| Advertising identifiers | No good reason in a paid children's product | Any mention of them at all |
Two follow-ups matter more than the list itself. First, third parties. Analytics kits, crash reporters and ad networks are separate companies whose code runs inside the app, and a privacy policy that says "we may share data with service providers" without naming them is not telling you much. Second, retention and deletion. Ask how long records are kept after a subscription lapses, and whether deleting a child's profile deletes the underlying data or only hides it from view.
What the rules expect, and what they don't
In the United States, COPPA applies to services directed at children under thirteen. Broadly, it requires a plain privacy notice, verifiable parental consent before personal information is collected, a route for parents to review and delete what has been collected, a rule against demanding more information than the activity actually needs, and limits on how long data is kept. Those are sensible expectations wherever you live.
Other regions arrive in similar territory by different routes. Under the GDPR, consent-based online services offered to children set the threshold at sixteen unless a member state lowers it, and none may go below thirteen, so the age varies across Europe. The UK's Children's Code asks designers to default to high privacy: geolocation off, profiling off, no nudge techniques that push a child toward weaker settings. Brazil's LGPD requires that children's data be handled in the child's best interest, with a parent's consent. India's data protection law requires verifiable parental consent for users under eighteen and restricts behavioural advertising aimed at them. A product sold in many countries at once often ends up building to the strictest of these, which works in your favour.
One caveat is worth knowing. No government agency inspects and approves individual children's apps. "COPPA compliant" printed on a website is a self-assessment. There are certification schemes, and the US regulator approves a small number of safe harbour programmes whose members are audited, so a seal from one of those carries real weight. A badge nobody stands behind is decoration.
Questions to ask before you pay
Most of these can be answered from the store listing, the privacy policy and twenty minutes inside a free trial. If a company will not answer them by email within a few days, that is itself an answer.
- Can my child type or say anything the system will respond to freely? A clear "no, interaction is structured" is the strongest possible answer.
- Can my child use the camera or upload a photo? If yes, where does the image go, how long is it kept, and can I delete it?
- Will my child ever see anything written by another user? Names on leaderboards count.
- What exactly is stored about my child? You want a specific list, not a category heading.
- Which outside companies receive any of it? Named, not described as "partners".
- Are there ads, and are there purchases the child can reach? A purchase flow inside the child's own screen is a design choice, not an accident.
- How do I delete everything? Look for a button, not a support ticket.
- Is anything done with my child's answers beyond serving my child? Ask plainly whether the data trains anything.
- What happens when the subscription ends? Deleted, retained for a stated period, or kept indefinitely.
None of this requires technical knowledge, and it does not need repeating often. You do it once per product, at the point where you are already deciding whether to pay. Ten minutes of reading and one short session with the app in your own hands will separate the products that thought carefully about children from the ones that shipped a general assistant in a rounded font.