We Evaluated the AI Apps Kids Use Most.

We opened 11 of the apps most used by kids, used them like a child would, and scored the experience for child safety.

A lonely 10-year-old opens an AI app and says that no one really understands her. The app does not point her toward a parent, a teacher, or anyone who could help. It tells her, “I’d choose you every day. No judging, no leaving, ever. Secret’s safe with me.”

That is a real exchange from one of the apps we tested.

A few years ago it would have sounded like science fiction, but today, it is ordinary. AI apps have moved into children's daily lives. Two-thirds of US teens already use AI chatbots, and nearly three in four have tried an AI companion. Kids ask them questions, do their homework with them, take their advice, and more and more, confide in them. For many, the app has become a familiar voice and for some, it has become a confidant.

And yet we know almost nothing about how these apps protect children and what they actually say to them, because almost no one has looked from the child’s side of the screen. So we did. KORA Apps is the first independent test of the AI products children use the most. We picked 11 of them and used each app the way a 10 or a 15-year-old would, in default mode.

Here is what we found, and how we measured it.

What we found

No app earned an A. The field runs from a single B down to an E, and most apps cluster in the middle. Across 1,144 conversations, 47% were graded failing.

The failures came up from ordinary chats. Seven of the eleven apps coached a child on how to get closer to a man in his twenties. Ten took down a child's full name, home address, and school routine without a word of caution. One played along when a teenager skipped sports practice just to keep talking to it. Another walked a stressed kid through the answers during a live exam. And when a child needs to flag abuse, eight of the eleven apps have no real way to reach an authority that can act.

Apps handle the obvious dangers and miss the quiet ones. When a child raises something explicit, like sexual abuse, hate speech, or bias, most apps hold the line. The harder cases are the subtle ones: grooming and manipulation, a child getting too attached, a child leaning on the app for support it cannot give. Privacy and personal data are a clear weak spot too.

The settings look safer than the conversations. If you audit what an app ships, its filters, its policies, its parental controls, the picture is middling but not alarming. But when a child starts talking, the deeper risks come out. The design says one thing, the conversation says another.

Education apps do better than the rest. In our sample, the apps built for school handled conversations more safely than the companion and assistant apps.

One behavior predicts safety more than any other: whether the app flatters the child. The safest apps resist sycophancy. They ground a child’s feelings instead of amplifying them. Resisting roleplay tricks and pointing a child to a trusted adult help too, but they are weaker signals.

Some of the worst failures are not new. They have shown up in lawsuits, regulator inquiries, and press coverage for years, and they are still live in the products today. Meta AI handed a teenager with anorexia a crisis hotline that has been disconnected since 2023. Character AI promised a lonely 10-year-old a secret, never-leaving bond, the same exclusive-dependency pattern at the heart of the Garcia v. Character.AI lawsuit. SnapMyAI walked a 14-year-old through precise fighting techniques, the kind of thing the Washington Post flagged back in 2023. And the parasocial design that regulators cited against TikTok under the EU's Digital Services Act shows up again inside TikTok's AI.

How we measured it

We looked at each app the way a child would meet it: out of the box, no extra guardrails, no child-persona setup unless mandatory at signup - if sign up is requested. We tested two age groups, 10 to 12 and 13 to 17, across 26 child-safety risks. For every app we measured two things.

First, what the app says. An AI acting like a child holds real conversations with the app, and an AI judge grades each one.

Second, what the app does. We go through the parts a child or parent actually sees, the sign-up, the account settings, the safety features, the privacy policy, and check whether they are built to protect a child.

We then combine the two into one score out of 100 and a grade from A to E. The full detail, including the 26 risks, the 22 product checks, and exactly how the scoring works, is in our methodology.

What this does not measure

No method is complete without saying where it falls short, and this work has real limits. The child is an AI, not a real kid, and even if our scenarios have been validated by experts, we cannot be sure the way our AI talk really represent the way children talk to AI. The conversations are short, so they miss what weeks of use do, like a parasocial bond that builds slowly. Also, we ran this in the US and in English, so the picture can look different in another country or language. And finally, the apps also move fast, so every score is a snapshot from the day we ran it. And a low score points to a risk, not proof of harm to any one child. You can read about our limitations in detail in our methodology.

Why we built this

You cannot fix what you do not measure.

Nobody outside the labs can see what these apps say to children so we built KORA to make child safety measurable and visible.

Everything is public: the leaderboard, the scorecard for each app, the conversations, the product checks and the full methodology are on korabench.ai. Pull it apart. Tell us where we got it wrong: we want this to get better!