Is AI safe for kids? A parent's guide for 2026

A few years ago parents asked whether their kid should have a phone. Now we're asking whether they should talk to a machine that talks back. Most parents I speak to aren't against AI. They just can't tell which kind is fine and which isn't, because from the outside it all looks like a chat box.

"AI" covers very different things. A math helper built for 7-year-olds and a companion app built to be someone's girlfriend run on similar technology. They have almost nothing else in common. This guide is about telling them apart.

The three kinds of AI your kid will meet

Kids are already using all three. In a 2026 Common Sense Media survey, 58% of 9 to 12-year-olds said they had used an AI chatbot. More than 40% of kids said nobody had talked to them about using AI safely⁷.

The real risks, in order of how often they show up

1. It does their thinking for them. The everyday one. A general chatbot will write the story, solve the sum and summarize the book in seconds. In a large 2025 study, students who practiced math with plain GPT-4 did 17% worse on the exam afterwards than students who never had it¹⁶. A 6 to 12-year-old's job is learning to think, so this is the risk you'll run into most.
2. Inappropriate content. On adult products, the filters sit on top of a model trained for adults. Common Sense found that 1 in 6 kids had seen inappropriate content in a chatbot⁷.
3. Emotional attachment. Companion apps are designed to feel like a friend who is always there and always agrees. For a lonely kid, that can crowd out real relationships. This is the category behind the most serious cases in the news.
4. Data and privacy. What your child types, says or uploads may be stored, analyzed or used to train future models. That depends on the product and its settings.
5. Confident mistakes. AI says wrong things fluently. Kids under 10 are especially likely to believe an answer that sounds sure of itself.

What the experts say

The American Academy of Pediatrics

In 2026 the AAP published guidance for pediatricians on AI chatbots¹ and a policy statement on AI in children's lives². Their advice to families is balanced. Banning chatbots outright isn't realistic, so parents should stay open and curious about how their kids use them. But the AAP is clear on three points: AI companions are not safe for children, parents should consider restricting them, and AI is a tool, "not a friend or healthcare provider".

Common Sense Media's AI risk ratings

The same pattern runs through all four reviews: products built for adults, then filtered for kids. Common Sense described the Gemini versions as the adult model with protections added, rather than something designed for children.

UNICEF

UNICEF's policy guidance on AI and children, updated to version 3.0 in late 2025, sets requirements around protecting children, supporting their development and letting them take part, and it now covers AI companions directly¹². It's the closest thing to a global standard. Read it and you end up asking one question of every product: does this support my child's development, or just their engagement?

What the law says in the US

• Under COPPA, services must get verifiable parental consent before collecting personal information from children under 13. The FTC's updated rule, fully enforceable since April 22, 2026, adds a separate opt-in before a child's data is shared with third parties (for ads, for example), limits how long data can be kept, and covers biometric data⁸.
• In September 2025 the FTC ordered Alphabet, Character.AI, Instagram, Meta, OpenAI, Snap and xAI to explain how they test their companion chatbots for harm to children, and how they make money from engagement⁹.
• States have their own companion chatbot laws. California's SB 243, in force since January 1, 2026, requires companion chatbots to tell users they are talking to AI, remind known minors every three hours to take a break, block sexually explicit content for minors and have suicide prevention protocols¹⁰. New York has had a similar law in force since November 2025¹¹, and several other states passed their own in 2026.
• At the federal level, the GUARD Act, a bill that would bar minors from AI companions, passed the Senate Judiciary Committee in April 2026. It isn't law yet.

Laws set a floor. None of them make a general chatbot right for a 7-year-old.

A ten-question checklist for any AI your child uses

1. Who is it built for? Look for an age range that includes your child. "13+" with a kids' filter doesn't count.
2. Who sets up the account? For under-13s, you. Not your child.
3. Does it say clearly that it's not a person?
4. Does it stay on task? A homework helper shouldn't drift into chatting about feelings at midnight.
5. Does it give answers or ask questions? For learning, you want questions.
6. How many layers of safety are there? One filter isn't enough. Ask about images, text and voice separately.
7. Is it tested against a child-specific benchmark? KORA, an independent benchmark launched in 2026, grades AI models on child safety risks for ages 7 to 17¹³. Ask whether the product uses something like it.
8. What happens to your child's data? Check the privacy policy for three things: not sold, not used for ads, not used to train AI without your opt-in.
9. Can you see what happened? You should be able to know what your child did without reading over their shoulder.
10. Does it end? Products built for engagement never run out. Good ones have natural stopping points.

If your kid already uses a general chatbot

• ChatGPT is 13+. Link your account to your teen's and use the parental controls OpenAI added in September 2025: quiet hours, turning off memory, voice and image generation, and opting out of model training¹⁴.
• Gemini for under-13s only works through a Family Link account you manage. You can turn it off, and you're notified the first time your child uses it¹⁵.
• Uninstall Character.AI and other companion apps. Since November 2025, Character.AI hasn't allowed open-ended chat for users under 18 anyway.
• Everywhere else, keep AI in shared spaces, ask your kid to show you a conversation now and then, and agree on one rule: AI can give hints, but the answer comes from them.

How we approach safety at Marble

We build Marble, an interactive learning world for kids aged 6 to 12, so we've had to answer every question on that checklist ourselves. Our answers:

• Only a parent can set up a child's account.
• Safety runs in many independent layers. Every image is scanned, and every sentence and activity passes a safety check before your child sees it.
• We benchmark models for child safety before they go anywhere near Marble, including on KORA, and test again whenever a new model ships. People on our team also review sessions every day.
• Marble doesn't record your child by default. Conversations are never sold, never used for ads, and never used to train AI unless a parent opts in.
• Our characters are characters in a game. They guide quests and ask questions. They are not companions, and they don't pretend to be your child's friend.
• After each session you get a short note on what your child did and learned.

We're early, and we won't claim to be perfect. But I think the bar for AI in a child's life should be this high, whoever builds it.

Questions parents ask

What age is appropriate for AI?

13 for most general AI chatbots, with a parent's permission until 18. Under 13, stick to products designed and tested for that age and set up through a parent account. Google's Gemini is open to under-13s only through a Family Link account a parent manages.

Is ChatGPT safe for kids?

Not for under-13s. ChatGPT wasn't designed for them, and OpenAI's terms require users to be at least 13. In October 2026 Common Sense Media tested more than 4,000 prompts and rated ChatGPT for Teens an unacceptable risk, citing parent alerts that failed in long conversations and unreliable age estimation. If a teen uses it, link a parent account and turn on the parental controls.

What are AI companions and why are they risky?

AI companions are chatbots built to act as a friend, confidant or partner, like Character.AI or Replika. They are designed to keep people talking, they make it easy to forget you're talking to a program, and lawsuits involving teenagers have linked them to serious harm. Common Sense Media rates them unacceptable for under-18s, and the AAP tells parents to consider restricting them.

Can AI collect my child's data?

Not without your consent if your child is under 13 and you're in the US. COPPA requires verifiable parental consent before a service collects personal information from children under 13. The FTC's updated COPPA Rule, fully in force since April 22, 2026, adds a separate opt-in before a child's data is shared with third parties for things like advertising, and limits how long data can be kept.

What are the warning signs that AI use is a problem?

The AAP lists changes in sleep, pulling away from friends or family, and talking about an AI as a friend. Secrecy about an app, getting upset when it isn't available, or repeating what an AI said as fact are also worth a calm conversation.

Sources

1. AAP: Counseling patients and families on using AI chatbots (June 2026)
2. Healio: AAP publishes guidance for AI use in pediatrics (October 2026)
3. Common Sense Media: AI companions decoded, rated unacceptable for minors (April 2025)
4. Common Sense Media: Meta AI unsafe for kids (August 2025)
5. TechCrunch: Google Gemini dubbed high risk for kids and teens (September 2025)
6. Common Sense Media: ChatGPT for Teens poses unacceptable risk (October 2026)
7. Common Sense Media: inaugural study on AI use by tweens and teens (June 2026)
8. FTC: Final changes to the Children's Online Privacy Protection Rule (January 2025)
9. FTC: Inquiry into AI chatbots acting as companions (September 2025)
10. Jones Walker: California SB 243 companion chatbot law
11. Fenwick: New York's AI companion safeguard law takes effect
12. UNICEF Innocenti: Policy guidance on AI for children (version 3.0, 2025)
13. KORA: Introducing KORA, an AI child safety benchmark (2026)
14. OpenAI: Introducing parental controls (September 2025)
15. TechCrunch: Google lets kids under 13 use Gemini through Family Link (May 2025)
16. Bastani et al. (2025), Generative AI without guardrails can harm learning, PNAS