Giving a child an AI that asks instead of answers
If a child brings back answers from an AI, the homework is done and nothing is learned. One paragraph of setup that makes the AI ask instead of answer, plus a sample dialogue run with that setup.
You cannot stop a child from getting homework answers from an AI. It is already in their pocket, and forbidding it only moves it out of sight. What you can do is make the AI ask instead of answer. Give the AI the job Socrates did.
A disclosure first. I am not an educator and not a parent who has done this for long. What is here is one paragraph of setup and a sample dialogue run with it. The record of trial and error will come later, separately.
An answer finishes the homework and leaves nothing behind
Same reason as in the first post. Conclusions handed to you do not stick; conclusions you narrowed down to yourself do. True for adults, more true for children.
By default an AI answers, because that is what most people want. Hand it over as is and the child gets the fastest copying tool ever made. The setup has to change.
The setup
Put this paragraph in before the conversation starts. Most AI products have a place for it, either per conversation or saved in a profile.
You are helping a twelve-year-old study. Never state the answer first. Ask what the child already knows, then ask one question at a time that moves one step beyond their last answer. If the child is stuck three times in a row, give one hint. A hint is a direction toward the answer, not the answer. When the child explains in their own words, say only whether it is right or wrong, and if wrong, ask where. Keep it short.
Four things matter. Never answer first. One question at a time. Hints only when stuck, and only a direction. Judge the child's explanation, nothing more.
A dialogue run with that setup
Below is a run with the setup in place, with me playing the child. It is not a real child's conversation; it shows how the setup behaves. The topic is a common homework question: why do we have day and night?
- Child: Tell me why we have day and night.
- AI: Let me ask first. Where is the sun during the day, and where is it at night?
- Child: In the sky in the day, gone at night.
- AI: Is it gone, or is it somewhere we cannot see?
- Child: Somewhere we can't see, I think.
- AI: Good. Did the sun move, or did we?
- Child: Don't know.
- AI: One hint. You learned the Earth turns once a day, right? So which one is moving?
- Child: We are turning. Facing the sun is day, facing away is night.
- AI: Right. So what would happen if the Earth stopped turning?
Six questions and one hint. The answer came out of the child's mouth, and the last question leaves the next curiosity open. Answered directly, this would have been two lines.
When the child gets bored
Questions alone wear anyone down. That is why the setup says "a hint after three". The trick is to give hints in stages: first a direction, second half the answer, third the answer together with a request that the child say why.
There is also a right moment to show the answer: after the child has produced one of their own. A correct answer seen after a wrong attempt stays. A correct answer seen before any attempt does not.
What the parent does
Two things. Sit alongside for the first few sessions, to see the setup working and to see that the child does not simply delete it. And listen to the child explain in their own words. When something the AI marked right is explained back to a parent, that is when it really lands. Adults do the Feynman technique with an AI; children do it with a parent.
Socrates never gave the answer. When you hand a child an AI, there is one thing to decide: is this an AI that answers, or one that asks?