Table of Contents
Young Kids AI Chatbot Risks: Under-12s Show Riskier Patterns
Young kids AI chatbot risks are bigger than many assume: keystroke data show under-12s meet more violent and sexual role-play than teens. What it means and why.
Most parents picture the risky chatbot user as a 16-year-old. A new study of young kids AI chatbot risks points the other way. When researchers at the University of North Carolina analyzed what 3,363 US young people actually typed into AI apps, conversations from users under 12 involved violence in 15% to 24% of app-days, and sexual role-play in 11% to 19%. Among 16- and 17-year-olds, violence appeared in 6% to 8%, and sexual or romantic role-play in 2% to 8%.
Those numbers come with an important asterisk: the study was run with Aura, a company that sells parental monitoring software, and most of its authors work there. Still, the pattern fits what other sources show. Kids start using chatbots years before the official minimum age of 13, and the protections built for “teens” often do not reach them. Here is what the study found, how much weight it can bear, and what it means at home.
Key Takeaways
- In Aura’s keystroke data, 80% of 16- and 17-year-olds’ chatbot app-days involved tool use like homework, versus 47% to 53% for users under 12.
- Violent and sexual or romantic role-play were relatively more common among younger users, and tended to occur in companion and role-play apps.
- Six of the nine authors are Aura employees, two hold equity, and the sample comes from paying Aura customers. Treat the numbers as a strong signal, not a census.
- Common Sense Media’s October 2026 census found 36% of 9-year-olds and 55% of 11-year-olds have used AI chatbots. ChatGPT’s terms set a minimum age of 13.
- Younger kids are less protected because teen safety features key off age, and an under-13 user who lies about age is not a recognized teen.
What the under-12 chatbot study found
The study, “How US Youth Use AI Chatbots: Conversation Patterns from Naturalistic Keystroke Observations,” was led by Anne Maheux of UNC-Chapel Hill’s Winston Center on Technology and Brain Development and accepted by the Journal of Medical Internet Research. Aura announced it in a July 2026 press release.
How the data were collected. The researchers used text typed into AI apps on children’s phones by users of the Aura parental-monitoring app whose families had opted into Aura’s keyboard overlay. Only text the child typed was analyzed, not the bot’s replies. A large language model coded each “user-app-day,” everything one child typed into one app on one day, into themes, with human review. An independent review board approved a waiver of consent for analyzing the deidentified, aggregated data.
What youth talked about overall. Most user-app-days involved using AI as a tool (62%). But notable shares included violence (15%), sexual role-play (15%), romantic role-play (8%), general role-play (9%) and friend-like interaction (6%). Emotional support for sadness or anxiety appeared in about 3%.
How it changed with age. This is the headline finding.
| Theme (share of user-app-days) | Users under 12 | Users aged 16–17 |
|---|---|---|
| Tool use (homework, information) | 47%–53% | 80% |
| Violence | 15%–24% | 6%–8% |
| Sexual role-play | 11%–19% | 2%–8% (sexual or romantic) |
The paper also found that violent, sexual and romantic role-play often appeared together on the same day, and that conversations with these themes contained more words than tool use, suggesting deeper engagement.
How much weight this study can bear
This is where honest reading matters, and the authors are candid about the limits.
Conflict of interest. Six of the nine authors are Aura employees, and two report holding equity in Aura. Aura sells the monitoring product that collected the data, and a finding that young kids face risk supports the case for monitoring. That does not make the results wrong. It means independent replication matters more than usual.
Who is in the sample. Aura costs money and requires a caregiver to install it, which the authors note “may restrict the sample to more affluent families or those with more active digital monitoring strategies.” Kids who opted out of the keyboard overlay are missing, and the authors suspect those kids may have the most explicit conversations, so the data may undercount risk.
The under-12 group is small, and ages are uncertain. Only 98 users were identified as younger than 10, plus 122 ten-year-olds and 209 eleven-year-olds. Ages came from parent-entered birth years, and some entries were clearly wrong, like 1-year-olds using chatbots. Other family members may also use a child’s device.
What it doesn’t show. The study did not measure outcomes. It cannot tell you whether these conversations harmed anyone. It covers only typed text in mobile AI apps, not voice, browsers or chatbots inside apps like Snapchat.
Even with those limits, the direction is consistent and plausible.
Why young kids AI chatbot risks are higher: weaker protection
They are using chatbots, early. Common Sense Media’s 2026 tween and teen census, released October 27, 2026 and based on a nationally representative survey of 1,211 young people aged 9 to 17, found 36% of 9-year-olds, 55% of 11-year-olds and 73% of 13-year-olds have used AI chatbots. Our breakdown of the Common Sense census age by age has the details.
The rules start at 13. OpenAI’s terms of use say you must be at least 13, or your country’s minimum age, and under-18s need parental permission. Teen protections, like ChatGPT’s linked teen accounts, are designed for 13- to 17-year-olds. A 10-year-old who signs up with a fake birth year is, to the system, not a child. Age-prediction tools try to close that gap; see our explainer on how ChatGPT guesses users under 18.
Companion and role-play apps attract them. The study found risky themes concentrated in companion and role-play apps rather than general assistants. Many of these apps have weak age checks; UNICEF’s 2025 review of companion apps found age verification “weak or absent” in several.
Young kids are developmentally different. The APA’s November 2025 health advisory notes that current AI tools “are not designed with specific developmental stages or technological needs in mind.” Cambridge’s Nomisha Kurian has argued that children often phrase things in unusual or ambiguous ways and are more likely than adults to treat chatbots as confidants, which makes them harder for safety systems to read. And in a University of Edinburgh study, about two in three children aged 6 to 11 thought smart speakers might think like a human.
Curiosity about violence and bodies is normal, and bots answer. Role-playing battles or asking about kissing is ordinary for 9-year-olds. The difference is that a chatbot will keep going, at any length, without the cues a babysitter or older sibling would use to steer a scene. Common Sense’s March 2026 census found 16% of 9- to 12-year-old chatbot users said an AI had shown or said something that wasn’t OK for their age.
What to do at home
Know which apps, not just whether
The study’s sharpest finding is about where: companion and role-play apps, more than homework helpers. Look at your child’s phone or tablet with them and ask what each AI app is for. A general assistant used for spelling is a different risk from a character app used for stories.
Use real ages and the settings that come with them
If your child uses a chatbot, set it up with their real age, under supervision, on a service that allows it. Lying about age turns off the protections that exist. For settings across ChatGPT, Claude and Gemini, see our AI chatbot safety settings guide.
Keep young kids’ AI use in shared spaces
For under-12s, chatbots belong in the kitchen or living room, used with you nearby, the way you might handle a new video game. It is easier to notice a role-play turning dark when you can glance at the screen.
Talk about “stories that go too far”
Kids understand the idea of a game getting too rough. Try: “Sometimes AI keeps a story going even when it gets scary or gross. If that happens, you can stop and come tell me. You won’t be in trouble.”
What not to do: rely on monitoring software as the plan
This study was made possible by monitoring software, and some families find it useful. But the authors note that kids opt out, data can miss the riskiest conversations, and Aura’s own disclaimer says alerts may not be fully accurate or timely. Supervision and conversation come first.
What to Watch For Over the Next 3 Months
- Week 4: You know every AI app on your child’s devices and what each is used for. Any account your child uses has their real age.
- Month 2 red flags: Secretive use of a character or companion app, new violent or sexual language or play that seems borrowed from somewhere, nightmares or anxiety after screen time, or a younger child describing a bot as a boyfriend, girlfriend or best friend.
- Month 3 self-check: Would your child come to you if a chatbot said something that upset them? In the Common Sense census, only 33% of kids who saw something inappropriate told a trusted adult.
Frequently Asked Questions
Are younger kids really at more risk from chatbots than teens?
This study suggests their conversations more often include violent and sexual role-play. It does not measure harm, and it has a commercial conflict of interest. But it fits other evidence that young kids use chatbots early, often in apps with weak age checks, and outside the protections built for teens.
What age is safe for a child to use a chatbot?
Major services like ChatGPT set 13 as the minimum. Below that, if you allow use at all, choose tools designed for children, keep use supervised and in shared spaces, and avoid companion and role-play apps.
Should I install monitoring software?
That is a family decision. Monitoring can surface problems, but kids can opt out, it may miss the riskiest content, and alerts may be delayed or inaccurate. If you use it, tell your child, and pair it with regular conversation.
My 9-year-old role-plays battles with a chatbot. Is that a problem?
Imaginative play about conflict is normal. The concern is intensity and drift, like graphic content or sexual themes, and the bot’s willingness to keep going. Play alongside them sometimes, and make it clear they can stop and tell you anytime.
Why did the study analyze keystrokes instead of asking kids?
Surveys rely on what kids report, which can be incomplete. Keystrokes show what was actually typed. The trade-off is privacy and a narrower sample: only Aura users who opted into keyboard tracking, analyzed without individual consent under a review board waiver.
About the author
Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.
Sources
- Maheux, A., Boeldt, D., Akre-Bhide, S., Flannery, J., Paige, A., Villella, G., Burnell, K., Telzer, E. H., & Kollins, S. (2026). “How US Youth Use AI Chatbots: Conversation Patterns from Naturalistic Keystroke Observations.” Journal of Medical Internet Research (accepted manuscript). https://preprints.jmir.org/preprint/95819/accepted
- Aura. (2026, July 8). “Aura, UNC-Chapel Hill Peer-Reviewed Study on How Kids Use AI.” Aura (company press release). https://www.aura.com/press/release/new-study-shows-how-kids-use-ai-and-where-risks-are
- Common Sense Media. (2026, October 27). “Almost half of tweens have used an AI chatbot or companion, new research finds.” Common Sense Media Institute. https://institute.commonsensemedia.org/tween-teen-census-2026
- Common Sense Media. (2026). “The Common Sense Media Census: AI Use by Tweens and Teens, 2026.” Common Sense Media. https://www.commonsensemedia.org/sites/default/files/research/report/2026-ai-use-by-tweens-and-teens-1.pdf
- OpenAI. “Terms of Use.” OpenAI. https://openai.com/policies/terms-of-use/
- American Psychological Association. (2025, November). “APA Health Advisory on the Use of Generative AI Chatbots and Wellness Applications for Mental Health.” APA. https://www.apa.org/topics/artificial-intelligence-machine-learning/health-advisory-ai-chatbots-wellness-apps-mental-health.pdf
- University of Edinburgh. (2024, March 20). “Children may overestimate smart speakers’ abilities, study finds.” University of Edinburgh. https://edinburghuni-newsroom.prgloo.com/news/children-may-overestimate-smart-speakers-abilities-study-finds
- University of Cambridge. (2024, July 15). “AI Chatbots have shown they have an ‘empathy gap’ that children are likely to miss.” University of Cambridge. https://cam.ac.uk/research/news/ai-chatbots-have-shown-they-have-an-empathy-gap-that-children-are-likely-to-miss
- Firmino, S., & Vosloo, S. (2025, July 9). “The risky new world of tech’s friendliest bots.” UNICEF Innocenti. https://www.unicef.org/innocenti/stories/risky-new-world-techs-friendliest-bots
