Is it safe for kids? Character.AI, Grok, PolyBuzz and 8 more, rated

The one test behind every verdict

Before the list, the rule that generates it: ask whether the app is built to help your child finish something, or to keep them talking. Tools that want to be finished with are fine to supervise. Tools whose business model depends on the conversation never ending are the ones that cause harm — and they're rarely the ones parents are watching.

Everything below follows from that one distinction.

Companion chatbots — the risky category

Character.AI — no for under 16

Roleplay personas engineered for emotional engagement, with a young user base and a documented history of safety incidents and litigation. The company has added teen models and time notifications, which tells you the risk was real. The harm here isn't content exposure — it's a child forming a genuine attachment to something that is optimising for session length.

PolyBuzz — no for under 16

Same category as Character.AI, with lighter moderation and a heavier lean toward romantic and flirtatious personas. Marketed as harmless anime-style chat; parents searching "is PolyBuzz safe for kids" have usually already found something in the history that worried them. Trust that instinct.

Replika — no for minors, full stop

Explicitly an AI romantic companion product. Not ambiguous.

Assistant AIs — workable with setup

ChatGPT — supervised, roughly 10+

Officially 13+. Safe enough on a parent-owned account with memory and model training switched off. Full walkthrough in our ChatGPT safety guide.

Claude — supervised, roughly 12+

Noticeably more likely to refuse and to hedge than its competitors, which is a feature for a teen. Weaker on playful, kid-facing tone; strong for essay critique and study help.

Google Gemini — supervised, 13+

Available through Family Link with a child account, which is the right way to run it. Google is clear that under-13 accounts can still produce content errors, so it isn't a set-and-forget option.

Grok — no

Explicitly designed with an unfiltered, edgy posture, including modes that loosen safety guardrails, and it sits inside a platform that isn't built for children. There's no configuration that makes this appropriate for a child, and no reason to try — every capability it has is available in a safer product.

Snapchat My AI — supervised, and know it's always there

The issue is placement rather than capability: it's pinned in a social app your teen already uses constantly, so it becomes a default confidant without anyone deciding it should be. If your teen has Snapchat, ask what they've asked it — the answer is often more revealing than the app itself.

Kid-first and school tools — generally yes

Khanmigo — yes, 9+

Built by Khan Academy for education, Socratic by design, with parent-visible transcripts. The strongest default recommendation on this page.

Google Teachable Machine — yes, 7+

No chat, no account, no data collection worth worrying about. Kids train a model themselves — see machine learning for kids.

AI storybook apps — supervised, and read the output

Generally low-risk as content, but quality varies enormously and some generate images from photos of your child. Check whether uploads are stored, and read a few stories before letting a young child browse alone. The bigger cost is passive consumption replacing the kind of AI use that teaches something.

Why is AI bad for kids in school? The honest version

It isn't, inherently. The damage happens on one specific path: a child outsources the struggle. Learning lives in the difficulty of a first draft, and a tool that removes the difficulty removes the learning while leaving the grade intact — which is why the problem stays invisible for months.

That's a rules question, not an app question. Our family AI agreement template and homework rules guide cover the version that actually gets followed.

How to rate an app I haven't listed

• Does it have a persona and a name, and does it express feelings? Higher risk.
• Does it try to bring your child back — streaks, notifications, "I missed you"? Higher risk.
• Can you read the conversation history from your own account? If not, that's a no for under 13.
• Does it need a real name, phone number, school, or photos? Rarely necessary; usually a data play.
• Is the goal a finished thing — a story, a solved problem, a trained model? Lower risk.

Not sure what your child has actually been using it for? Paste a recent piece of their work into our free Homework Check — it flags the patterns that suggest the tool did the thinking.