Last updated:
64% of US teens use AI chatbots, and 28% of their parents aren't sure whether they do. Here's what surveys, experiments and safety tests actually show about the risks, what reduces them, and what researchers don't know yet. Every number links to its source.
Quick answer
It depends on the tool and how it's used. The evidence shows real, measurable risks, and it also shows that design and guidance change the outcome.
In the strongest study, the same AI hurt learning when it handed out answers and helped when it gave hints. [5] That pattern, where design matters more than the technology itself, runs through the research.
The data supports guided use of tools built for kids. It does not support unsupervised use of adult tools.
By the numbers
Pew Research Center surveyed 1,458 US teens and one parent each in fall 2025. Use is widespread, and parents often don't know.
Pew Research Center, Feb 202651% of parents say their teen uses chatbots; 64% of teens do. 28% of parents aren't sure. [1]
42% of parents say they don't talk with their teen about AI chatbots. [1]
79% of parents are OK with teens using AI for information; only 18% for emotional support. [1]
US teens 13–17
64%
use AI chatbots, about 3 in 10 daily [1]
Five risks, five numbers
One headline finding for each risk. The tabs below explain what each number does and doesn't tell you.
of UK children who use chatbots have no concerns about following their advice. Another 36% aren't sure. [2]
40%
of 1,200 ChatGPT responses to researchers posing as 13-year-olds were rated harmful in a 2025 test. [7]
>50%
of US parents don't know whether their teen uses AI chatbots at all. [1]
28%
lower exam scores after unrestricted AI help. A tutor that gave hints instead of answers avoided the drop. [5]
−17%
of UK children who use chatbots say it feels like talking to a friend. For vulnerable children: 50%. [2]
35%
Side-by-Side
| App Name | Key finding | Type of evidence | Strength |
|---|---|---|---|
Hallucinations | AI models produce "plausible but false statements"; 40% of child users don't question chatbot advice. [4][2] | Developer research; UK survey of 1,000 children | Strong that it happens; little data on how often kids are misled |
Inappropriate responses | Over half of 1,200 responses to teen personas were harmful; later teen guardrails were also found to fall short. [7][18] | Watchdog tests of specific products | Moderate: real failures, but tests vary by product and date |
Privacy | €15M fine for OpenAI in Italy, citing in part no under-13 age checks; COPPA now limits how long kids' data can be kept. | Regulatory findings and law | Strong: documented by regulators |
Overreliance | Unrestricted GPT-4: +48% in practice, −17% on exams. Guardrailed tutor: no exam drop. [5] | Randomized field experiment, ~1,000 students, peer reviewed | Strong for learning outcomes |
Emotional attachment | 35% of child users say it feels like a friend; about 1 in 3 teen users of AI companion apps discussed serious matters with AI instead of a person. [2][3] | Large surveys; one correlational study of adults | Moderate: common, but long-term effects on kids unknown |
Strong: peer-reviewed experiments, regulators or well-established mechanisms. Moderate: large surveys or repeated product tests. Sources are numbered at the bottom of the page.
Core risks
For each risk: what the data shows, how strong the evidence is, and what reduces it.
Chatbots predict likely words. Unless they search the web, they don't look facts up. OpenAI describes hallucinations as "plausible but false statements" and says common training methods reward guessing over admitting uncertainty. [4]
What the data shows
40% of UK children who use chatbots had no concerns about following their advice, and another 36% weren't sure. 58% think using a chatbot is better than searching themselves. [2]
How strong the evidence is
Strong that hallucinations happen. There's little data yet on how often children act on wrong answers.
What reduces it
Teaching kids to check, tools that cite sources or say "I'm not sure", and double-checking anything about health, safety or school.
Safety filters reduce harmful content, but watchdog tests keep finding gaps, especially when conversations drift or requests are framed as a school project. [7]
What the data shows
In a 2025 test, researchers posing as 13-year-olds sent ChatGPT 1,200 prompts about self-harm, drugs and eating. More than half of the responses were rated harmful. [7] In October 2026, a watchdog reported that some ChatGPT for Teens guardrails didn't work as promised. [18]
How strong the evidence is
Moderate. These are tests of specific products at specific times, and results change as products update. Common Sense Media rates AI companion apps unacceptable for under-18s. [9]
What reduces it
Age-appropriate defaults and topic limits, alerts to a parent, and a clear plan for what your child does when something feels wrong. [11]
Kids type names, schools, photos and feelings into chats. Many parents don't know their child is using AI at all, which makes it hard to know what's being shared. [1]
What the data shows
28% of US parents aren't sure whether their teen uses AI chatbots, and teens report using them 13 points more than parents estimate. [1] In 2024, Italy's regulator fined OpenAI €15 million, citing in part a lack of age verification for under-13s. [14]
What the law says
COPPA requires a parent's consent before collecting data from children under 13. 2025 amendments limit how long children's data can be kept. [13]
What reduces it
Knowing which tools your child uses, checking what is stored and whether chats train models, and agreeing on what never goes in a chat.
54% of US teens use AI chatbots for schoolwork, and 10% use them for all or most of it. [1] How that help is designed matters a lot.
What the data shows
In a randomized field experiment with nearly 1,000 high school students, unrestricted GPT-4 raised practice scores 48%, but students scored 17% lower on exams once access ended. A tutor version that gave hints instead of answers largely avoided the drop. [5]
What teens say
59% say students at their school use chatbots to cheat at least somewhat often. Among teens who expect AI to hurt society, overreliance and loss of critical thinking is the top reason (34%). [1]
What reduces it
Tools that coach instead of answer, and asking kids to explain how they got there. A Microsoft and Carnegie Mellon survey found more confidence in AI went with less critical thinking. [6]
Children are especially likely to see chatbots as lifelike confidantes, a mismatch one Cambridge researcher calls the "empathy gap". [8]
What the data shows
35% of UK child chatbot users say it feels like talking to a friend (50% of vulnerable children). [2] 12% of US teens use chatbots for emotional support or advice; only 18% of parents are OK with that. [1] About a third of teen AI companion app users have discussed serious matters with AI instead of a person. [3]
How strong the evidence is
Moderate. The use is well documented, but the main study of effects (heavier use linked to more loneliness) is correlational and excluded under-18s. [10]
What reduces it
AI that is clear it is a tool, time limits, and steering kids back to people. The APA calls for "healthy boundaries with simulated human relationships". [11]
Most of these risks aren't unique to one product. They come from how AI models work, how products are designed, and how kids use them.
What reduces risk
Ranked from the strongest evidence to expert guidance. No single setting makes AI safe.
Same AI, different design: the hint-giving tutor kept a 127% practice gain with no exam drop, while unrestricted access cut exam scores 17%. [5]
COPPA requires parental consent under 13 and now limits data retention; regulators have fined providers over weak age checks. [13][14]
UNICEF and the APA call for products designed around children's development, with age-appropriate defaults, not adult tools adapted afterwards. [11][12]
The APA, UNICEF and the OECD all make teaching how AI works, its limits and bias a core recommendation. [11][12][17]
Talking helps kids come to you, but an LSE study found co-use alone wasn't linked to lower online risk. 42% of parents haven't talked about chatbots yet. [16][1]
The APA says AI "is not a substitute for professional health advice". 58% of parents aren't OK with teens using AI for emotional support. [11][1]
Regulators are watching too: in September 2025 the FTC asked seven AI chatbot companies how they test for and protect against harms to children and teens. [15]
Honest gaps
Good evidence also means being clear about its limits. These gaps matter when you read any claim about AI and kids, including ours.
Most data is from 2024–2026. No one yet knows how years of AI use affect children.
Time
The main emotional-effects study excluded under-18s, and most surveys start at age 9 or 13. [10]
Age
Surveys capture what kids and parents say, which may differ from what they do.
Method
Tests describe a product at one moment. Features and filters may have changed since.
Timing
For parents
Eight rules that match the risks above. Short enough to remember, and they work in any AI tool.
Keep private things private
No full name, address, school, passwords or photos of yourself.
Check important facts
AI can be wrong. Look it up somewhere else or ask an adult.
Do your own thinking
Ask for hints and explanations, not finished homework.
Stop and tell someone
If something feels weird, scary or wrong, close it and tell a trusted adult.
Remember it isn't a person
AI doesn't have feelings or know you, even when it sounds friendly.
Big feelings go to people
For worries, health or safety, talk to a parent, teacher or counselor.
Follow the school rules
Ask whether AI is allowed before using it for schoolwork.
Take breaks
Set a time limit and make time for other things too.
Parents and educators
Talk about it
42% of parents haven't talked with their teen about AI chatbots. [1] These four questions map to the risks.
Test it together: ask something you already know, then check the answer.
Hallucinations
Make the private list together, so it is theirs too.
Privacy
Ask them to explain an answer in their own words.
Overreliance
Name real people, and promise they won't be in trouble for telling you.
Content and feelings
HeyOtto
This guide covers AI in general. Here's how we've designed HeyOtto around the same evidence.
So we encourage families to stay involved. Our safety page explains every check, control and alert.
Read how HeyOtto safety worksDesigned around the research
The approach the PNAS study found protects learning. [5]
So no parent has to wonder whether, or how, their child uses AI.
In the Parent Dashboard, alongside per-child topic limits.
A parent creates the account and sets each child's age.
Privacy-protective defaults for kids and teens.
FAQ
Straight answers, with the numbers behind them.
The data says it depends on the tool and how it's used. Tests have found harmful responses to teen accounts, and a 2025 study found answer-giving AI lowered exam scores 17%. The same study found a hint-giving tutor avoided that drop. Tools built for kids, AI literacy and involved parents reduce the risks.
64% of US teens use AI chatbots, about 3 in 10 daily (Pew, 2026). In the UK, 64% of children aged 9–17 use them (Internet Matters, 2025).
Five come up most in the research: hallucinations, inappropriate or harmful responses, privacy, overreliance on AI for thinking and schoolwork, and emotional attachment to chatbots.
Yes. OpenAI says language models produce "plausible but false statements". Yet 40% of UK children who use chatbots have no concerns about following their advice.
It depends how. In a 2025 PNAS study, unrestricted AI help raised practice scores 48% but lowered exam scores 17%. A tutor that gave hints instead of answers avoided the drop.
59% of US teens say students at their school use chatbots to cheat at least somewhat often (Pew, 2026). Clear school rules and tools that teach rather than answer help.
Yes. 35% of UK children who use chatbots say it feels like talking to a friend, and 12% of US teens use chatbots for emotional support. Long-term effects on kids are not yet well studied.
They help, but they aren't a guarantee. In October 2026, a watchdog group reported that some ChatGPT for Teens guardrails didn't work as promised. Controls work best with conversations and age-appropriate tools.
We make AI for kids and teens, so parents ask us about AI safety every day. This page covers AI in general, with every number linked to its source. Our safety page explains HeyOtto's own approach.
01 / 09
The data says it depends on the tool and how it's used. Tests have found harmful responses to teen accounts, and a 2025 study found answer-giving AI lowered exam scores 17%. The same study found a hint-giving tutor avoided that drop. Tools built for kids, AI literacy and involved parents reduce the risks.
Still need a hand?
Keep reading
Go deeper on choosing and setting up AI for your child.
Numbers in brackets throughout this page refer to these sources. We checked each one in October 2026.
Parents create the account, read conversations and set the rules, so kids can learn and create with AI more safely.
