High School Students Who Used AI for Math Practice Scored Worse When It Was Taken Away. Here Is the One Skill Worth Protecting Before That Happens to Your Kid.

My youngest had a homework assignment last week that involved explaining how something worked in science. She opened AI, read the explanation it gave her, nodded to herself, and typed a version of it into her document. She did not struggle. She did not have to. I watched the whole thing happen in about four minutes. I did not say anything, because I could not name exactly what bothered me about it. A new report in Science News gave me the language.

Research now shows that when AI replaces cognitive work entirely, the skill being replaced quietly shrinks, even in people who were already highly competent.

A team from the Wharton School at the University of Pennsylvania ran a study where high school students used ChatGPT to work through math practice problems.^[1] Their scores improved while they had access to the tool. That part matches what most parents have heard about AI in education. But when the AI was removed and the students were tested without it, they performed worse than students who had never used ChatGPT at all. The tool had not been teaching them math. It had been doing math for them, and their own ability had drifted in the process.

A second study, published in The Lancet, followed gastroenterologists who used an AI detection tool during colonoscopy procedures for three months.^[2] Their detection accuracy went up while they had help. When the tool was removed, their rates of finding polyps dropped below where they had started before using AI. These were trained doctors with years of developed skill. Three months of AI reliance was enough to reduce their independent accuracy. The researchers called it a deskilling risk.

A third study from Carnegie Mellon found that students who used AI assistance to practice SAT reading comprehension questions struggled to answer questions independently when the tool was gone.^[3] They were also more likely to skip questions than students who had never used AI for practice. Researchers described what was happening as a use-it-or-lose-it effect: when a skill stops being practiced, it fades the way a muscle does inside a cast. The AI had not been a learning aid. It had been a substitute for the cognitive work that builds capability.

Your kid is starting a school year right now, and the first weeks tend to set patterns that persist through the rest of the year.

Most conversations about AI in education focus on cheating, or on whether a particular tool is allowed at school. The research points to a different problem, one that does not require any bad intent to produce real harm. A student who uses AI to get answers genuinely feels like they are working. The answers appear. They write them down. They move on. Nothing signals that the skill is not forming. The gap is invisible until the AI is gone and the test is in front of them.

My middle daughter is nine and learns best when she is enjoying herself. She would use AI for everything if I let her, not because she wants to avoid thinking, but because AI is patient and always gives her something. What the research suggests is that the something it gives her, without requiring her brain to work for it, may be exactly the problem. She feels like she learned. The capability is not actually building.

The PNAS math study is worth sitting with because the students involved were not trying to cheat. They were using AI the way teachers and parents often encourage: as a learning support. And it still reduced their independent ability once the tool was removed. That finding changes the question from “are they misusing AI?” to “how are they using it?” The distinction between those two questions is where the skill lives.

The skill is Active Thinking: keeping your brain engaged in the cognitive work even when AI is available to do it for you.

The same research that found AI causes skill loss also found a way to prevent it. A companion study from Wharton tested students who received hints from AI rather than complete answers, and found their skills actually improved compared to students who had no AI at all.^[4] The difference was not the technology. It was whether the student’s brain stayed in the loop. Hints require effort to use. Answers do not. That gap in required effort is where the capability either forms or does not.

Benjamin Lira, a behavioral scientist at Wharton involved in the coaching study, offered a phrase I have not stopped thinking about since I read it: outsource the crap, not the craft.^[4] The research shows that using AI to handle mechanical, routine parts of a task, like formatting a document, checking spelling, or finding a definition, does not cause skill loss. Outsourcing the part that requires genuine thinking, the analysis, the argument, the original problem-solving, is where the atrophy happens. Most kids have never been taught to tell those two things apart.

Trent Cash, a behavioral scientist at the University of Waterloo, put it plainly in an interview with Science News: “If you’re not engaging with the cognitive work, you’re not going to learn.”^[5] Active Thinking is not about avoiding AI or limiting screen time. It is about one specific thing: staying mentally engaged in the part of the work that builds the skill. That engagement can happen alongside AI. It just has to happen.

A child building Active Thinking can tell you why an answer is right, not just what the answer is.

One clear sign is explanation. If your kid used AI for something and you ask them to walk you through how it works, and they can do it in their own words without needing to open the AI again, their brain was in the loop while they were using it. If they look back at the AI and describe what it said rather than what they understood, that is a signal the cognitive work was being skipped. The explaining is not a quiz. It is a check on whether the understanding actually formed.

A less obvious sign is how they respond to difficulty. My oldest, who is thirteen and extremely goal-focused, has an instinct I have noticed since she was younger: when something is hard, she treats that as information about what she needs to practice, not as a reason to find a shortcut. She does not panic when she gets stuck. She adjusts. In an AI world, that instinct, staying with the difficult part rather than routing around it, is exactly what prevents the skill loss the research describes. It is not the most visible habit to build, but it is the one that matters most when the AI is not there.

After the next AI session, ask your child one question: can you explain that in your own words?

Make explaining part of the routine. After any homework session where AI was involved, take two minutes and ask your kid to walk you through one thing they learned, without opening the AI again. You do not need to evaluate whether the answer is right. You are checking whether they can put it in their own words. If they can, the cognitive work happened. If they cannot, you have a natural opening to work through it together. One question, done consistently, builds the habit of checking your own understanding before closing the tab.

Try the five-minute rule before opening AI. Before your kid asks AI a question for school, encourage them to try on their own for five minutes first. Even an unsuccessful attempt matters. The research on hints versus answers suggests the effort of trying, getting stuck, and then using AI to get unstuck produces better learning than going straight to AI. The five minutes of trying is what keeps the skill forming. You do not need to enforce it strictly. Suggest it once and see what happens over the next few weeks.

Help name what to outsource and what to own. Spend five minutes together going through a recent assignment and asking: which parts did AI do that your brain should have done? Formatting, spell-checking, and definitions are generally safe to hand off. The thinking, the analysis, and the original argument belong to your kid. That conversation, done once, gives them a framework they can apply on their own. Most kids have never been asked to make this distinction, and once they have, they notice it themselves.

The good news from this research is that the situation is not inevitable. Skill atrophy happens when AI replaces thinking. It does not happen, and can even reverse, when AI coaches instead. That gap is a choice, and it is one a kid can make once they know the difference exists.

I started paying closer attention to how my youngest uses AI after reading this. What I noticed is that she is fast. She finds answers quickly, which means she accepts them quickly and moves on. Nothing in the experience of using AI tells her that the speed is the problem. Active Thinking is not about slowing down. It is about one moment, after the AI gives its answer, where she asks herself: do I actually understand this, or did I just watch it happen? That question is the whole skill. It takes about four seconds to ask. But whether she asks it is the difference between using AI and learning from it.