Your Kid Is Using ChatGPT for Homework. Here's What That's Actually Doing to Their Brain.

My son came home last Thursday with an A on his history essay. I should have been thrilled. Instead, I opened his Google Docs and checked the version history.

He'd typed four words into the document himself. Four. The rest appeared in a single paste at 9:47 PM, eleven minutes before the deadline. When I asked him about the thesis — his thesis, the one earning him an A — he couldn't explain it. He looked at me the way you look at someone who just asked you to recite a phone number you never memorized.

That was the night I stopped worrying about whether ChatGPT would replace teachers and started worrying about what it was doing to my kid's ability to think.

I spent the next two weeks reading everything I could find. What I learned scared me more than any social media study ever has.

The numbers are worse than you think

Between May and December 2025, the percentage of middle school, high school, and college students using AI for homework jumped from 48% to 62%, according to a RAND Corporation study published in March 2026. That's not a slow creep. That's a majority of students, in seven months, deciding that the way they've always done schoolwork is obsolete.

And here's the part that should alarm you: the students themselves are worried. The same RAND study found that more students now believe AI is eroding critical-thinking skills than believed it a year ago. They can feel it happening. They're using it anyway.

Pew Research Center data shows the number of teens aged 13 to 17 using ChatGPT for schoolwork doubled between 2023 and 2024 — from 13% to 26%. By now, that number is almost certainly higher, because OpenAI made ChatGPT free for anyone with an email address and loosened the signup requirements to a point where age verification is effectively nonexistent for anyone who can type a birthdate.

The crutch study that should be on every parent's fridge

A research team at the University of Pennsylvania's Wharton School ran what might be the most important education experiment of the decade. They gave nearly 1,000 high school students in Turkey access to GPT-4 as a math tutor, then measured what happened when they took it away.

The results, published in the Proceedings of the National Academy of Sciences, should make every parent pause.

During the AI-assisted practice sessions, students with unrestricted GPT-4 access performed 48% better than the control group. Students with a specially designed tutoring interface performed 127% better. The numbers look incredible until you read the next paragraph.

On the follow-up exam — with no AI — the students who'd had unrestricted access scored 17% worse than peers who never used it at all.

Let me say that differently, because the implication matters. The kids who practiced without AI outperformed the kids who practiced with it, once the AI was gone. The GPT-4 group didn't just fail to learn; they learned less than if they'd struggled on their own. The researchers' term for it was "crutch" — students offloaded the cognitive work to the machine and never built the neural pathways that practice is supposed to build.

There's a silver lining buried in the data, and I'll get to it. But first, parents need to understand why this isn't just a homework problem.

What's happening inside their brains

An MIT study — covered by TIME in early 2026 — tracked brain activity while participants wrote essays under three conditions: no AI, Google search only, and ChatGPT. The ChatGPT group showed the lowest brain engagement across every measure. Neural activity dropped. Linguistic complexity dropped. And here's the detail that made me put down my coffee: the ChatGPT users got lazier with each successive essay. By the final round, most had stopped writing entirely and were copy-pasting.

That pattern — where the tool makes the first attempt easier, which makes you lean on it harder the second time, which makes you incapable of the third attempt without it — isn't new. Psychologists call it cognitive offloading, and it's well-documented in calculator dependence and GPS navigation studies. What's new is the speed. A calculator offloads arithmetic. ChatGPT offloads reasoning itself. The entire process of organizing thoughts, evaluating evidence, constructing an argument, recognizing when your argument has a hole — that's what gets outsourced.

A scoping review in Computers & Education Open cataloged the risks across K-12 settings and identified three damage categories. The first is psychological: emotional disconnection from learning and social isolation when schoolwork becomes a solitary transaction between a kid and a chatbot. The second is intellectual agency: reduced independent learning, inflated self-assessment (kids think they understand the material because they got the right answer, even though the AI did the understanding), and weakened creative ownership. The third is systemic: schools don't have governance frameworks, teachers don't have training, and the equity gap widens because wealthier kids get supervised AI use while lower-income kids get unsupervised dependence.

Actually, let me revise that third point — the equity concern cuts both ways. Wealthier families also produce more unsupervised AI users, because the kids have their own devices in their rooms with no one watching. The assumption that affluence equals oversight is wrong, and I've seen it in my own home.

What OpenAI says — and what they don't

OpenAI requires users to be at least 13. Between 13 and 18, parental consent is technically required, but the verification mechanism is a checkbox. In January 2026, OpenAI introduced parental controls — you can now disable voice mode, image generation, and memory features for family accounts, and set quiet hours when ChatGPT can't be used.

These are welcome steps. They are also not the point.

The point is that no parental control setting addresses the learning damage. You can lock down every feature and your kid can still paste their essay prompt into the chat window, receive a fully formed argument they didn't construct, and submit it. The tool works exactly as designed. The problem isn't misuse. The problem is that the core use case — "help me with my homework" — is the one causing harm when there are no guardrails on how that help is delivered.

This is where the Wharton study's silver lining matters. The researchers found that when GPT-4 was redesigned as a Socratic tutor — giving hints aligned with the teacher's pedagogy instead of direct answers — the learning damage disappeared. Students using the guided version actually retained what they learned. The AI asked questions back. It pointed students toward the right method without doing the method for them.

The difference between "here's the answer" and "what do you think the next step is?" turned out to be the difference between learning and not learning.

What you can actually do tonight

I'm not going to tell you to ban ChatGPT. Your kid would find a workaround before you finished reading this sentence, and the 32% age-check bypass rate from the Internet Matters UK report applies to every content restriction, not just social media. Bans don't work. Supervised use does.

Here's what the research actually supports:

Sit with your kid the first three or four times they use ChatGPT for schoolwork. Not hovering — sitting. Watch what they type. Ask them what they're going to do with the answer before they read it. This isn't surveillance; it's the same thing you did when you taught them to cross the street. The Children and Screens research institute specifically recommends this approach in their 2026 parent guide.

Teach them the "explain it back" rule. If they can't explain the AI's output in their own words, they didn't learn it. This is dead simple and works at every age. My son hated it the first week. By the third week, he started asking ChatGPT different questions — not "write me an essay about the Industrial Revolution" but "what were the three biggest arguments against child labor laws in 1830s England?" He was using it as a research tool instead of a ghostwriter. The quality of his thinking changed visibly.

Consider tools designed for kids. Khan Academy's Khanmigo uses the Socratic questioning approach from the Wharton study — it won't give your kid the answer. It asks them what they think first, then guides them through the reasoning. It's not perfect, and it's not free for all features, but it's built on the principle that the struggle is the learning.

For younger kids — elementary and middle school — the research is unambiguous: they should not be using generative AI for schoolwork at all. The foundational skills that AI bypasses (reading comprehension, mathematical reasoning, written expression) are still being built. Offloading them at this stage isn't a shortcut; it's skipping a developmental phase that doesn't come back.

The conversation that matters more than the settings

My wife and I sat down with both our kids last month and said something I never thought I'd say to a twelve-year-old and a fourteen-year-old: "This tool is going to be part of your life forever. We need to teach you how to use it without it making you dumber."

That word — dumber — landed hard. It was supposed to. Because every piece of evidence I've read says the same thing: unsupervised, unrestricted AI use for schoolwork makes kids less capable of independent thought, less accurate in self-assessment, and less equipped for the exact future they'll need to navigate. The American Prospect published an investigation in February 2026 calling it "the unregulated AI experiment on America's children." That framing isn't dramatic. It's descriptive.

I don't know where this ends up. The Wharton study proves that AI tutoring can work when it's designed to protect learning, but most kids aren't using Khanmigo. They're using ChatGPT, which gives them whatever they ask for, instantly, with no friction. And friction — the productive struggle of figuring something out — turns out to be the entire mechanism through which learning happens.

My son still uses ChatGPT. He uses it differently now. He used it last Tuesday to generate counterarguments to his own thesis, then spent forty-five minutes dismantling them. His teacher marked it as his best critical-thinking work of the semester.

That's the version of this technology I want him to know. It took me watching, questioning, and being annoying about it for three weeks before he got there on his own.

Nobody said parenting in 2026 would be easy. But this one isn't optional.

Digipalz teaches kids and families the digital safety skills schools don't cover — from AI literacy to social media resilience. See how our program works →

If your child is struggling with anxiety, depression, or thoughts of self-harm related to academic pressure, contact the 988 Suicide & Crisis Lifeline by calling or texting 988.