AI can produce the thing that makes it look as though we did them. Google could find the information. AI can read it, compare it, summarize it and produce the answer we once took as evidence that somebody understood it.
AI can outsource the evidence that thinking occurred.
That could be wonderful. A child with a good idea may no longer need years of technical training before turning it into a design, program or invention. Henry may someday point a camera at a broken toy and ask what is wrong. AI could identify the motor, explain the wiring and show him how to replace it.
The trouble is that the finished product may mature faster than the child. A ten-year-old may produce work that looks adult without understanding much of what made it good. The familiar worry is cheating, but children were avoiding homework long before ChatGPT. The deeper problem is that AI may remove the clumsy stages through which people become competent.
We learn to write by producing sentences that deserve a quiet burial. We learn research by following promising sources into swamps. We learn cooking by burning dinner and mechanical work by discovering that a screw can, in fact, be stripped beyond recognition. Most of all, we learn judgment by being wrong while the consequences are still small.
AI may eliminate drudgery. Unfortunately, some of the drudgery was where we hid the education.
In a randomized experiment involving nearly a thousand high-school math students, those given unrestricted AI assistance did better while they had it—and worse than the control group when it disappeared. Students using a version designed to give hints rather than answers largely avoided that loss.³ The same technology can be a tutor or a crutch.
Schools are going to have to stop asking only, “Did you use AI?” and start asking, “What did the AI do, and what did you still have to do?” Did it explain the hard part and challenge an assumption, or did it supply the conclusion, write the argument and leave the child with the exhausting responsibility of changing the font?
Banning AI would prepare GenAI for a world that will not exist. Letting AI do everything would be like giving children a chauffeur and calling it driver education.
The better model may already be on Henry’s floor. Lego instructions are useful because they teach structure and sequence. By following them, Henry learns enough to depart from them. AI should do the same. It should help children understand the pieces, not deliver the finished model hot-glued inside the box.
The purpose of education is not merely to produce competent output. It is to develop the judgment and curiosity to know what to do when the instructions run out.
There is another danger we should see coming. AI is cheap to scale; human attention is not. That creates an obvious temptation for schools with the fewest resources to use AI not alongside teachers, tutors and counselors but in place of some of them.⁴
One group gets AI plus people. The other risks getting AI in place of people.
A chatbot can answer thirty children at once without demanding health insurance or complaining about the school budget. It may not notice that one child has stopped asking questions, another is frightened and a third has just discovered something worth encouraging. AI with a teacher could expand opportunity. AI instead of a teacher could become cheaper neglect with excellent grammar.
For now, Henry turns the broken toy over, studies the wheels and looks for the battery compartment. More important, he notices that another child has a problem he might be able to solve. AI may someday help him diagnose it, but it did not give him the curiosity, the pleasure of making something work or the impulse to help the child holding it.
Hunter watches. He sees Henry follow the pictures, find the pieces and build what the instructions show. Then he sees him push the booklet aside and create something nobody showed him.
A machine may someday answer every question Hunter asks. It cannot be his older brother.