Teen AI use: what the data actually says.

The strongest response to children’s AI use is neither panic nor unrestricted access. It is informed supervision: knowing what kind of tool a child is using, what job they are asking it to do, what information they are sharing, and who remains accountable for the result.

The headline numbers

Pew Research Center surveyed 1,458 U.S. teens ages 13–17 and their parents from September 25 to October 9, 2025. The survey was designed to represent U.S. teens in that age range who live with a parent. Its overall margin of error was plus or minus 3.3 percentage points.

Those percentages do not mean every use was frequent, unsafe, or academically inappropriate. “Ever used” includes a wide range of experiences. Pew also found that about three in ten teens used chatbots daily, while other teens used them less often or not at all.

Schoolwork: help, overreliance, and authorship

Pew found that one in ten U.S. teens said chatbots help with all or most of their schoolwork. Another 21% said chatbots help with some of it, and 23% said a little. Forty-five percent said they had not used chatbots in this way.

Research help and math support were among the more common school uses. Editing was less common, although 35% of teens said they had used a chatbot to edit something they wrote. At the same time, 59% thought students at their school use AI to cheat at least sometimes.

A practical authorship test

Before submitting AI-assisted work, a learner should be able to answer four questions:

• What part of the work did I create?
• What part did the tool suggest or change?
• Which claims did I verify, and where?
• Could I explain every idea in my own words without the chatbot open?

If the answer to the last question is no, the tool may have replaced learning rather than supported it.

AI assistants and AI companions are not the same thing

An AI assistant is usually used to answer a question or complete a task. An AI companion is designed—or used—to create conversation that feels personal, relational, or emotionally meaningful. The categories can overlap, but the risks and family rules should not be treated as identical.

Common Sense Media surveyed a nationally representative sample of 1,060 U.S. teens ages 13–17 in April and May 2025. Its definition focused specifically on AI companions, not ordinary homework helpers, image generators, or voice assistants.

The same report found that 33% of companion users had chosen the AI instead of a person for an important or serious conversation at least once. But the picture was not uniformly negative: 80% of companion users said they spent more time with real friends than with AI companions. Research should be read as a pattern of benefits and concerns, not proof that every user is isolated or harmed.

What the risks actually are

1. Confident misinformation

A language model can produce invented citations, incorrect calculations, or incomplete advice in fluent language. The problem is not simply that it can be wrong; it is that the tone may not reveal the uncertainty.

2. Privacy without a clear boundary

Children may treat a chat box like a private conversation and share names, locations, school information, health concerns, images, or another person’s story. The service’s storage, training, moderation, and deletion practices vary by provider.

3. Overreliance and reduced ownership

When a tool supplies the structure, wording, and conclusions of a task, a learner can finish without practising the underlying skill. Overreliance is about the role the tool plays—not merely how often it is opened.

4. Bias and missing perspectives

Outputs can reflect gaps or stereotypes in training data, evaluation, and product design. Children need practice asking whose perspective is missing and whether the answer would change for another person or context.

5. Relationship-like design

A chatbot’s warmth, agreement, availability, and memory can make it feel more understanding than it is. It does not owe a duty of care, keep professional confidentiality, or reliably recognize when a child needs human support.

UNICEF’s 2025 child-centred AI guidance emphasizes safety, privacy, fairness, transparency, wellbeing, inclusion, and preparing children for an AI-shaped world. Its 2026 multi-country snapshot also cautions that evidence is still sparse and adoption is moving faster than research.

A proportionate family response

Rules work better when they describe a decision, not just a banned product. A family can use the following ladder:

Five family rules worth writing down

1. We do not enter passwords, precise location, private images, health details, or another person’s information.
2. We check important claims with a reliable source or qualified person.
3. We follow the teacher’s AI policy and disclose help when required.
4. We do not treat friendly language as proof that a system understands or cares.
5. We can show an uncomfortable conversation to a trusted adult without getting in trouble for asking for help.

How to read the research responsibly

These studies answer different questions. Pew measured broad chatbot use; Common Sense Media focused on AI companions; UNICEF combined evidence from multiple countries and projects. Their percentages should not be added together or treated as directly comparable.

Survey answers are self-reported. Definitions, products, safety features, and teen behaviour change quickly. A percentage can describe prevalence, but it does not prove that AI caused a particular outcome. It also does not tell a parent what one individual child is doing.

The most useful conclusion is therefore modest: children are encountering AI at meaningful scale, the context of use matters, and families need practical literacy and open communication more than a single universal rule.