Why kids who argue with AI are building something the others aren't
MIT put 54 university students into three groups: ChatGPT users, search engine users, no tools. EEG caps throughout. By the third essay, the ChatGPT group had stopped writing and were asking the AI to do it. The implication for children is uncomfortable.
In 2025, researchers at MIT's Media Lab ran 54 university students through three groups: one wrote essays using ChatGPT, one used a search engine, and one had no tools at all. The participants wore EEG caps throughout, and brain activity was measured across 32 regions.
The ChatGPT group showed the lowest neural connectivity by some distance. By the third essay, several of them had effectively stopped writing and were just asking ChatGPT to produce the work for them, and two English teachers who later read the final essays described them as soulless. The lead researcher, Nataliya Kosmyna, named the effect cognitive debt.
The participants were adults, which makes the implication for children uncomfortable to sit with.
What the brain does when it accepts an answer
Critical thinking, in this context, is the difference between accepting an answer and engaging with it. Acceptance is essentially passive, while engagement is active, and the brain does almost no work in the first mode and a great deal in the second.
When you accept an answer your job ends, but when you engage you have to think the problem through yourself, hold the AI's answer alongside your own version, find the gap, and pick which is right. That second process is where reasoning gets built.
Why this isn't just an MIT story
A 2025 paper in Societies, surveying 666 people, found that frequent AI use correlated with lower critical thinking scores, with the youngest participants leaning hardest on the tool. A 2025 randomised controlled trial added a different angle: students given unrestricted ChatGPT access during their study sessions retained noticeably less than the students who studied without it.
None of this is an argument against AI; it's an argument about when in the process AI should show up. The flipside is also documented: a month of coding can change how six-year-olds plan, inhibit impulses, and reason through problems that have nothing to do with a screen. Both effects come from the same place. What the brain does when you make it do the work itself.
What this looks like at the kitchen table
Ask the AI a question, read the answer out loud, then turn to your child and ask: what would you say back?
The phrasing matters. "Is the AI right?" is a yes-or-no question and it usually gets a quick yes, whereas "what would you say back?" puts the child in the conversation rather than on the receiving end of one. They can't answer without thinking the problem through.
A nine-year-old asked why ice melts will probably say something half-right, while the AI's answer will be tidier and more polished. Compare the two and find the gap, because the gap is where the thinking happens. Twenty minutes of an eight-year-old working something out for themselves, with paper and pennies and damp socks, builds more than ten polished answers handed over.
What the best teachers have known for thirty years
Robin Alexander and Neil Mercer at the Cambridge Faculty of Education have spent decades on something called dialogic teaching, where lessons are built around children questioning, challenging, and reasoning out loud rather than being told. The evidence is robust: children taught this way learn more, and the effect holds across subjects.
AI raises the stakes on getting this right, because a classroom full of children who have spent five years passively accepting AI answers won't suddenly start engaging when the GCSEs arrive.
Three habits that build the reflex
Use "what would you say back?" after any AI answer that matters, because the question turns reception into engagement in a single sentence.
Once a week, ask the AI something your child already knows well and let them correct what comes back, because the role reversal builds the muscle better than any explanation could.
The third habit is a homework rule. AI now does plenty of things children used to do alone, like maths problems, short essays, and comprehension answers, and the rule is that AI can be a starting point but never a finishing point. Whatever it produces, the child has to add to, change, or argue with before it counts as theirs.
This one is more demanding for the parents than for the children. Pushback is contagious, but only if a parent does it first, and adults who outsource their own thinking to AI can't reasonably expect their children not to. The other two pieces in this series take different angles on the same instinct: the trust calibration that comes before pushback, and the verification habit that comes after it.
