Last updated September 20, 2026, with a correction to how ChatGPT's teen account is switched on.
No. The largest study yet, the OECD's PISA 2025, found no evidence that AI makes children less clever. It found something narrower: the students who hand over the work score lower, while students who use AI to help them learn score about the same as those who never touch it. These are associations, not proof, and nobody has tested this on younger children. Tonight, check the clock, not the tool.
What the evidence says, in four lines
Is AI making children less clever? No, and none of the research below measures intelligence. It measures what a child can do on a test with no help, which is a different thing and the thing exams ask for.
Then what did the biggest study find? The OECD's PISA 2025, published September 8, 2026 from more than 760,000 fifteen year olds across 91 countries, found that students who use AI for specific schoolwork tasks score lower, while weekly users of AI for general learning score about the same as non-users.
Which children actually lose out? The ones whose homework got faster while their marks went up. In a study of 26,811 Chinese secondary students, the losses sat almost entirely with that group, and the students whose homework still took about as long lost almost nothing.
What should I do tonight? Nothing dramatic, and nothing that needs an app. Ask how long the homework took, then ask your child to talk you through one paragraph of it out loud.
What the OECD actually found
It found that the gap depends on the job the child gives the machine. PISA 2025, published on September 8, 2026, tested more than 760,000 fifteen year olds across 91 countries and economies. Its executive summary puts the AI finding in one sentence:
"Students who do not use AI for specific schoolwork tasks outperform those who do, while weekly users and non-users of AI for general purposes perform similarly."
Read that twice, because almost every headline carried the first half and dropped the second.
The specific schoolwork tasks are the ones you would guess: drafting a piece of writing, summarizing a set reading, doing the first sweep of research on a new topic. Those are the jobs where the students who hand over the work fall behind. The gap is not small. In the report's figures, as reported on September 9, 2026, fifteen year olds who almost never used AI for writing assignments averaged 509 points in science, against 481 for daily users, a gap of 28 points once family background is accounted for, which the OECD puts at roughly a year and a half of instruction.
Using it to learn is a different picture. There, weekly users and non-users land in much the same place.
The gap the OECD found is not between children who use AI and children who do not. It is between children who use it to finish and children who use it to learn.
Two more things from the same report matter at home. About one student in five now uses AI for schoolwork almost every day, and 14 percent say they never or almost never do, so this is already the normal condition of secondary school rather than a coming problem. And the OECD notes that students "tend to perform slightly better when frequent AI use is paired with opportunities to develop AI literacy skills", while warning that those opportunities "are more common among advantaged students, risking a growing AI divide."
That last sentence is the one I would not let go past. The protection is a classroom habit, and it is not being handed out evenly.
Did AI cause the fall in PISA scores?
Almost certainly not, and it is worth saying plainly because the two stories broke on the same morning. PISA 2025 recorded the lowest average scores the OECD has ever measured in mathematics and reading: down 14 points in reading and 9 in mathematics against 2022, with one in five fifteen year olds now a low performer in all three subjects, up from 16 percent. The OECD does not attribute that decline to AI, and the decline was already running before chatbots reached most classrooms.
The AI finding is a pattern inside the data, not the explanation for it. The OECD is careful to call these associations, and so am I: students who lean on AI to produce work may also be students who were already struggling, or less interested, or less supported at home. The report cannot separate those. What it can do is show you which uses travel with lower scores, consistently, across 91 school systems.
The study that watched the same children before and after
For cause rather than correlation, the strongest evidence is smaller and comes from China. David Strömberg, Victor Lei and Yanhui Wu tracked 26,811 students in grades 7 to 12, roughly 12 to 18 years old, across 30 months, and published it as a CEPR discussion paper on June 2, 2026. Rather than surveying students, they compared the same schools before and after generative AI arrived.
After it arrived:
- Homework scores rose about 18 percent.
- Time spent on homework fell about 30 percent, reported by Al Jazeera on September 2, 2026 as an average drop from around 64 minutes to around 45.
- Monthly closed-book exam scores fell about 20 percent within six months.
- Scores on the high-stakes entrance exam fell 18 to 24 percent, with the full effect arriving after roughly two years.
That is the frightening half. Here is the half the coverage keeps leaving out. The authors found the losses sat almost entirely with the students whose homework times collapsed while their homework marks climbed, which they call homework outsourcing, about 80 percent of the AI users. The students whose homework still took about as long as it used to lost almost nothing.
Same chatbot, same assignment, two children. The one who used it to get unstuck kept the learning. The one who used it to get finished did not.
One caveat, and it is a real one. This is a single country and a school system built around one enormous exam. The design beats a survey, because it follows the same students through the change, but it is not a randomized trial and it cannot tell you what will happen at your kitchen table.
Is this about intelligence, or about effort?
Effort, and one study isolates it cleanly. In the 2023 to 2024 school year, Hamsa Bastani, Osbert Bastani, Alp Sungu and colleagues ran a randomized trial with nearly 1,000 students in grades 9 to 11 at a large high school in Turkey. Some worked through practice problems with a standard ChatGPT interface. Some used a version built to tutor, giving hints and asking questions instead of handing over answers. Some had neither.
The standard interface lifted practice scores by 48 percent. Then the researchers took it away and ran an exam. Those students scored 17 percent below the students who had never had it.
The tutoring version lifted practice scores by 127 percent, more than twice as much. On the exam it did no measurable harm at all.
Same model, same students, same subject, opposite result. The variable was not the technology. That difference is the whole of the test I would run before paying for one: whether the tutor will withhold the answer.
Where did the "55 percent less brain activity" number come from?
From a small study of adults that has been retold until it stopped being about adults. The figure circulating in parenting posts traces back to work from the MIT Media Lab, posted by Nataliya Kosmyna and colleagues in June 2025. Fifty-four people wrote essays with a chatbot, with a search engine, or with neither, while their brain activity was recorded. The chatbot group showed the weakest neural connectivity and the lowest sense of owning what they had written. Only 18 of them carried on into the fourth session, where the groups swapped tools.
The participants were adults. And when I traced the "55 percent" version to its most shared article, it named no study, no sample size and no date, only a podcast segment.
In July 2026, Jorge Pereira Campos and Tatiana Koff wrote in Frontiers in Developmental Psychology that the adolescent is missing from this whole field. Their name for the risk is the Cognitive Offloading Paradox: AI is most tempting for exactly the mental jobs, planning, weighing evidence, building an argument, that a teenager has to practice in order to become able to do them. Their verdict on the evidence is blunter: the field has studied adults, they write, "yet the population with the most at stake has been largely absent from this work", and the evidence that would settle the worry "does not exist yet".
Something real is happening. Nobody has yet measured it in children.
What children say about it themselves
They say they can feel it. In Teens in the AI Era, published on August 18, 2026, Common Sense Media surveyed a nationally representative sample of more than 1,000 US teenagers. Seventy percent use AI for schoolwork. Among those who do, 25 percent use the answer as it arrives, 31 percent rewrite it to sound like their own, and 35 percent improve it with something they know.
And 39 percent say they feel like they are missing out on learning when they use AI to complete assignments.
That is the figure I would put on the fridge. Nearly four in ten teenagers have already noticed. The conversation at home does not have to open with an accusation. It can open with a question they have already asked themselves.
What I See in Schools
The pattern I meet most often is not cheating. It is work that got faster and flatter. Teachers describe pieces that are correct, tidy, well organized and completely silent about the child who handed them in. No wrong turn, no odd example, no sentence only that student would write. The grade is fine. The person is missing.
The schools handling this best did not buy better detection software. They moved the thinking back into the room: the first draft written in class, the two-minute check where a student explains one paragraph out loud, the closing question that asks what you changed and why. None of that is anti-AI. It is assessment that notices whether anyone learned anything.
And the OECD has now put a number on the other half of that work. Among students using AI every day, science scores ran 13 points higher, more than half a year of teaching, where lessons regularly asked them to judge the quality of what an AI had produced. Around six in ten students said they get that kind of lesson. The advantaged ones get it more often.
The question I hear at almost every parents' evening, or parent-teacher conference, is some version of: how do I know if she is actually learning it? No app will tell you. But there are two things you can see, one at home and one in the school's answer.
The Clock Test: the thing you can see at home
The Clock Test is the one check I would run on AI and homework at home: has the work got faster without getting harder? If a task that used to take forty minutes now takes eight and the grades went up, the effort moved to the machine, and the learning went with it.
Time is the only part of this you can watch without reading a word of your child's homework or installing anything. It is also, as it turns out, the exact signal the CEPR researchers used to separate the students who lost learning from the students who did not.
The Clock Test is not a verdict. Fast is sometimes just fluent, and a child who has genuinely got the hang of quadratics should be finishing quicker. Use it to start a question, never to land an accusation.
Two older ideas from this blog do the rest. The Homework Test is one question to ask of any piece of homework before AI touches it: if the machine did this part, what would my child lose? If the answer is nothing, let it help. If the answer is the learning, that part stays with your child. And underneath all of it: let AI help with the doing; keep the thinking with your child.
If you want three things to look at rather than one, look at what your child produced, how they produced it, and what they can now do without help. Only the third one turns up in an exam hall.
What to ask the school
Three questions, and none of them is "do you allow AI?", because that has only a policy for an answer.
- The Judging Question: how often is my child asked, in a lesson, to judge something an AI produced and say why it is wrong? This is the one classroom habit the international data links to higher scores among children who use AI every day, and roughly four in ten students are not getting it.
- Where in my child's subjects does work still get done without help, and how often? A child who writes in class regularly cannot quietly stop practicing.
- What happens if a teacher believes AI did the work? You want to hear about a conversation and a piece of evidence, not a percentage from a detector. If that has already happened in your house, what to do when your child is accused of using AI walks through the conversation.
For the wider version of this conversation, what your child's teachers are actually doing with AI covers where it sits in a school, where help ends and cheating begins covers the rules, and how to teach your child to question an AI answer is the home version of the habit the OECD data points at.
What to do this week, by age
Under 8. This is not really a homework question yet. OpenAI's own help center states that ChatGPT is not intended for children under 13, and that where under-13s use it in education an adult must be directly involved. The habit worth building here is doing the first try together, and explaining what AI actually is in words a young child can hold on to. If reading is the work at this age, the mechanical part of it, turning letters into sounds, only improves when the sound comes out of your child, out loud, to a person.
You could say: "Let's have a go with our own brains first, then we can ask the computer if we got it right."
8 to 12. Build the first attempt offline. Not because AI is dangerous at this age, but because the first attempt is where the learning lives, and the habit has to exist before the pressure arrives.
You could say: "Show me the bit you worked out yourself. That's the part I actually want to see."
13 to 15. This is where the Clock Test earns its keep, and where the OECD's split is most useful: ask what the tool did, not whether it was used. If your teenager uses ChatGPT, ChatGPT for Teens began rolling out to eligible teen accounts globally from August 18, 2026, with availability varying by region. OpenAI's help center states that you do not have to turn it on: where an account is eligible and its age information, verified age or age prediction indicates someone under 18, ChatGPT enables it automatically. It carries a study mode offering hints, step-by-step guidance, follow-up questions and knowledge checks rather than finished answers. What is not automatic is when that mode runs. Study hours, which OpenAI says a teenager or a linked parent can set, decide when eligible new chats open in it, through the parental controls setup described elsewhere on this blog. That is closer to the tutoring version in the Turkey trial than to a plain chatbot, though nobody has tested it the way those researchers tested theirs.
You could say: "I'm not going to ban it. I want to know which part of this you'd struggle with if it disappeared, because that's the part you're being graded on."
16 to 18. Exams are the whole game now, and the entrance-exam finding from the China study belongs here: the gap appeared years after the homework started looking better. In England, Wales and Northern Ireland this is a rules question too. JCQ's guidance on AI use in assessments, current version April 2025, allows AI in the research and preparation stages of coursework and non-exam assessment, but the student must name the tool and the date, keep a non-editable copy of the prompt and the output, such as a screenshot, and submit it with a brief explanation of how it was used. Students cannot be credited for work that is not their own, and unacknowledged use is malpractice, with sanctions up to disqualification.
You could say: "If this tool vanished the week before your exam, what would you be shaky on? Let's start there."
What none of this can guarantee
No setting proves a child thought. Study Mode, school filters, a family rule, this article: none of them can tell you the effort stayed with your child. OpenAI says plainly that its parental controls do not let a parent read or monitor a teenager's conversations. Only your child's explanation of their own work can do that, which is why two minutes of "talk me through this paragraph" beats any control panel.
School rules come first. If your child's school has an AI policy, or an exam board rule like JCQ's applies, that outranks anything you set up at home, and what a school ban does and does not change is worth reading before you assume the policy is doing the work. If your child has an education plan that includes assistive technology, whether a 504 plan, an IEP or an EHCP, the plan comes first and the tools in it are there for a reason.
And if the worry underneath this is your child's learning rather than AI, an AI rule will not fix it. That is a conversation with the class teacher, and sometimes a specialist, and it should not wait for a technology debate to finish.
The reason to stay calm is not that the risk is imaginary. It is that the risk turned out to be specific, and specific risks can be managed. If you are measuring this against the last technology you worried about, the social media playbook only half transfers: the talking half does, the restricting half does not. Your child does not need to use less AI. They need to keep doing the hard part first.
If your child's school has not spoken to parents about how AI is changing what gets assessed, that conversation is overdue, and it is the work I do with schools: training teachers on AI in the classroom so the policy means something in the room. It is worth forwarding to a head of year. I also write a newsletter that keeps teachers and parents up to date on this without the panic, and parents are very welcome.
Dan Fitzpatrick is a former secondary teacher and assistant headteacher, a bestselling author on AI in education, and the founder of The AI Educator. More about Dan.
