How much should teenagers use AI?

Nobody who tells you a number of hours is working from evidence. There is no study that produces one. There is one experiment that matters more than every survey put together, and it says something more useful: the same tool raised students' marks by 48 per cent while they had it and left them 17 per cent worse than students who never had it once it was taken away. A differently configured version of the same tool did not do that. The variable was not how much. It was how.

This page is written for you rather than about you, which mostly means two things. It does not assume you are in trouble. And it shows you the problems in the numbers adults are quoting at you, because you are entitled to check them.

The experiment that should change how you use it

Bastani and colleagues ran a field experiment with nearly 1,000 high-school students, split three ways: unrestricted access to GPT-4, a purpose-built tutor that gave hints but would not hand over solutions, and a control group with neither.

While the tools were available, both AI groups did better. Grades rose 48 per cent with unrestricted access and 127 per cent with the guardrailed tutor.

Then the researchers took the tools away and tested everyone again. The unrestricted group scored 17 per cent lower than students who had never had access at all. The guardrailed tutor group largely did not show that penalty.

Sit with the shape of that. Both groups looked like they were learning. Their marks went up. One group was learning and one group was borrowing, and no measurement taken during the term could tell them apart. The difference only appeared in the exam room with the tool gone, which is the only place it was ever going to appear.

The caveat the authors put on it: this was school mathematics over a bounded period, and it does not tell you what a good guardrailed interface looks like for other subjects or for professional work.

The practical version reduces to one question, the same one this site puts to chief executives: could you still do it without the tool, and when did you last check? Not a rule about hours. A test you can run on yourself, described in full at am I becoming dependent on AI.

The 72 per cent you keep hearing about

Common Sense Media surveyed 1,060 US teenagers aged 13 to 17 between 30 April and 14 May 2025, weighted to be representative, margin of error plus or minus 4.2 percentage points. The headline: 72 per cent have used an AI companion at least once and 52 per cent use one at least a few times a month.

Now read the definition the survey gave people before asking. It listed Character.AI and Replika, then added that it could also include using sites like ChatGPT or Claude as companions, even though those tools may not have been designed to be companions.

So a teenager who has ever had a rambling conversation with ChatGPT is inside the 72 per cent. The report says so itself in its limitations: some respondents may have conflated general AI use with AI companion interactions, potentially inflating usage statistics. That sentence is on page 12. It is not in the press release.

Here is the rest of the same survey, which travels much less well.

• 80 per cent of users spend more time with real friends than with AI companions, 68 per cent of them much more. 6 per cent spend more time with AI.
• 67 per cent find AI conversations less satisfying than conversations with real friends, 47 per cent much less satisfying.
• 50 per cent distrust the advice AI companions give. Only 23 per cent trust it quite a bit or completely.
• 74 per cent have never shared personal or private information with one. 66 per cent have never felt uncomfortable.
• 9 per cent regard an AI companion as a friend or best friend. 6 per cent say it helps them feel less lonely.
• 46 per cent describe them as tools or programs. The top two reasons for use are that it is entertaining (30 per cent) and curiosity about the technology (28 per cent).

The report's own discussion says most teens approach these tools pragmatically rather than as substitutes for human relationships. That is a fair reading of its data. It is not the reading that reached you.

The British survey, and the thing wrong with both of them

Internet Matters surveyed 1,000 UK children aged 9 to 17 and 2,000 parents in April and May 2025, with four focus groups of 27 teenagers and 17 days of testing three chatbots using fictional child profiles. It found 64 per cent of children had used an AI chatbot: ChatGPT 43 per cent, Google Gemini 32 per cent, Snapchat's My AI 31 per cent.

What they use them for, among users: schoolwork 42 per cent, finding information 40 per cent, curiosity 40 per cent, chatting 24 per cent, seeking advice 23 per cent, fun or escapism 18 per cent. Wanting a friend came in at 6 per cent and emotional help or therapy at 3 per cent, the two lowest categories on the list.

The finding that deserves attention is about a specific group. The report classifies children as vulnerable if they have an Education, Health and Care Plan, receive SEN support, or have a physical or mental health condition requiring professional help. Among that group, 71 per cent use chatbots against 62 per cent of their peers, they are nearly three times as likely to use companion-style products (17 against 6 per cent), 50 per cent say it feels like talking to a friend, and 23 per cent say they use one because they have no one else to speak to, against 12 per cent overall.

That is a serious finding and this page does not soften it. It is also where the number problems start.

Two reports, two internal contradictions, two press releases that are worse

You are old enough to be given the working, so here it is.

• Internet Matters describes two charts as having the same base, children who have used at least one chatbot. One uses 133 vulnerable and 499 non-vulnerable respondents. The other uses 188 and 802. Those cannot both be right, and 990 is close to the whole sample of 1,000 rather than the 638 who are chatbot users. The companionship figures, including the 50 per cent and the 23 per cent above, sit on that chart. The report does not flag the discrepancy or publish any significance testing.
• Its limitations section covers only the user testing. Nothing is said about the survey, the focus groups, or the size of the vulnerable subgroup, which is somewhere between 133 and 188 people.
• Its press release swaps the base of a figure. The report says 47 per cent of 15 to 17 year olds used a chatbot for homework. The release runs 42 per cent, which is the all-ages number, under the same claim.
• Common Sense Media contradicts itself by six points. Its topline data gives the share of users who transferred social skills to real life as 33 per cent. Its report body and chart give 39 per cent. Both documents are published.
• Its press release asserts something the survey cannot support. The release says the technology is already impacting teens' social development and real-world socialisation. The survey is explicitly cross-sectional and contains no causal or longitudinal measure. The release also names teens with mental health difficulties as especially vulnerable, and no figure in the report breaks results down by mental health status.

None of that makes either report worthless. Both are the best available data and both are more careful than their own publicity. It does mean that when someone quotes one of these numbers at you as settled fact, the correct response is to ask what the base was.

The question this page will not answer

Whether it is bad to talk to an AI when you are lonely sits underneath a lot of this. It goes unanswered here.

The reason is that the evidence does not exist yet. Everything available is cross-sectional: it can show that lonelier young people use these tools more, and it cannot show which came first. Publishing a confident answer would mean inventing one, and the subject is serious enough that inventing one would be worse than saying nothing. This research keeps that question on a watch list and will write it when there is something to write.

What can be said without any data at all: if you are using something because there is nobody else, the problem being described is the absence of the somebody, and that is worth telling one actual person about.

What actually follows from the evidence

Six things, each traceable to something above rather than to adult instinct.

• Use the version that makes you work. The guardrailed tutor beat unrestricted access on the measure that counted afterwards. If a tool will only give you hints, that is the feature and not the limitation.
• Test yourself with it switched off, on purpose, before something matters. The unrestricted group in Bastani had no idea until the exam. You can find out earlier, cheaply, any time you like.
• Write your own answer first, then ask. You cannot tell where a machine's answer diverges from a position you never formed. The argument is at human at the start.
• Assume the confident tone means nothing. Fluency is a property of the writing rather than of the knowledge, set out at why AI sounds so confident. Sixteen and seventeen year olds in the Internet Matters focus groups had already worked this out unprompted.
• Notice the difference between using it and talking to it. Most teenage use in both surveys is homework, information and curiosity. That is not the thing anyone is worried about, and conflating the two is how the 72 per cent got made.
• Ask for the base. Every number on this page has one and every base is printed. Numbers that arrive without one are being used on you rather than shown to you.

And one thing for the adults reading over the shoulder

The frequency question is the wrong one and it is the one every parental control is built around. Bastani's two AI groups would have looked identical on any usage dashboard. The measurable difference was in the design of the tool and it only surfaced when the tool was removed.

A household rule about hours regulates the variable that did not matter in the only experiment that has tested this properly. A household habit of occasionally doing something unaided regulates the one that did. UNICEF's guidance for AI systems affecting children is normative rather than empirical. Read it next anyway: it makes no claims about learning effects and does not pretend to, which puts it ahead of most of what is written for parents.

Key research and primary sources

• Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakci, O. and Mariman, R. (2025). Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematics.
• Robb, M. B. and Mann, S. (2025). Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions. Common Sense Media.
• Internet Matters (2025). Me, myself and AI: Understanding and safeguarding children's use of AI chatbots.
• UNICEF Innocenti (2025). Guidance on AI and Children, Version 3.0.
• Fisher, M., Goddu, M. K. and Keil, F. C. (2015). Searching for explanations: How the Internet inflates estimates of internal knowledge. Journal of Experimental Psychology: General, 144(3).

Both survey reports were read in full, including their toplines and methodology sections, and the discrepancies described above were found by comparing the published documents against each other and against their own press releases.

Related SuperSkills research

On younger children, should children use AI. On school and assessment, assessing students when AI can do the assignment and does AI detection work. On subject choice, what should I tell my children to study. On the self-test, am I becoming dependent on AI. On the everyday questions next door, letting AI summarise what you read, whether you still need to remember things and AI for therapy or advice.

About this research

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026) and founder of The SuperSkills Intelligence Company. Both survey reports and both press releases were read at source and every figure checked against the published document rather than against coverage of it. Where a report and its own topline data disagree, both numbers are given. The question of whether it is harmful to talk to an AI when lonely is deliberately not answered here, because the available evidence is cross-sectional and cannot separate cause from correlation. Not clinical advice. Reviewed quarterly.