When AI helps kids learn, and when it quietly doesn't

Two kids, one worksheet, very different weeks

Picture two children stuck on the same page of maths homework. One opens a chatbot, types in the question, and copies down whatever it hands back. The page is finished in ten minutes and looks perfect. The other uses a tool that flatly refuses to just tell her, and instead nudges her through it one step at a time, so the same page takes half an hour and she still gets two of them wrong.

Which child learned more? The obvious guess is wrong, and there is now solid evidence for why.

The study that should give parents pause

A 2025 study by researchers at the University of Pennsylvania put nearly a thousand high school maths students in Turkey through a term with AI help, split three ways. One group got a plain ChatGPT-style assistant. One got a tutored version built to offer hints and guidance instead of finished answers. The third had no AI at all, just the usual textbooks.

During practice, the AI groups looked brilliant. The plain-assistant group scored 48% higher than the no-AI group on the practice problems, and the guided-tutor group scored 127% higher. If you had stopped measuring there, you would have called AI a runaway success.

Then the researchers took the AI away and set everyone an exam to sit on their own. The plain-assistant group scored 17% worse than the children who had never touched AI. Not level with them. Worse. The tutor group, the one that only ever got hints, came out about even with the control group: no exam boost, but no damage either.

Hamsa Bastani, who led the work, put the lesson plainly. “If we use it sort of lazily and kind of outsource the work that we’re supposed to be doing and completely trust the machine learning model, then that’s when we could be in trouble.”

It isn’t one odd result

You could wave away a single study. This is not a single study. In 2026, a separate team led by Kamron Rismanchian sifted through millions of interactions on ALEKS, a maths platform used in schools from fifth grade up to college, comparing the years before and after ChatGPT arrived. Once the chatbot was widely available, high school students spent about 31% less time on word problems. And on supervised placement tests, where no AI was allowed, the share of word problems answered correctly slid from roughly 80% down to about 60%.

Graphing problems, the sort you cannot easily hand to a chatbot, barely moved. It was specifically the work children offloaded that they quietly stopped being able to do. One of the researchers called it “cognitive surrender,” and worried it was not going to stay confined to maths.

The hopeful half of the picture

Here is the part that keeps this from being a doom story. The very same technology, pointed differently, genuinely helps.

Take Rori, an AI maths tutor built by an education network and delivered over WhatsApp, of all things, so that a child needs only a cheap phone and a signal. It does not sit there dispensing answers. It walks a student through a problem and coaches the thinking, with small lessons about sticking with a hard question and treating mistakes as normal. In a randomised trial, children using Rori outscored their peers in maths by 11 percentage points. It has now been used by more than 100,000 students across Africa.

Same underlying models. Opposite result. The difference was never the AI itself. It was whether the tool did the child’s thinking for them, or made them do it.

What to actually look for

Full disclosure: we build Socratic AI tutors for children, so we came to this evidence with a view already formed. But that view is precisely why we build them the way we do, and none of the studies above are ours.

So when your child reaches for AI on their schoolwork, the useful question is not “is this cheating” or even “is this allowed”. It is simpler than that. Watch what the tool does the moment your child is stuck. Does it hand over a finished answer, or does it ask a question back and make them take the next step themselves? A tool that just answers is a very fast way to finish tonight’s homework and a slow way to get worse at the subject. A tool that makes your child do the reaching, and is willing to leave them stuck for a minute, is doing something close to what a good teacher does.

The homework getting done was never the point. The child being able to do the next one, without the phone in their hand, is.