Rule 1 — What never goes into the box
Children are used to search engines, where typing something has no consequence. A chat interface feels like a conversation, and children disclose to conversations.
The rule is a flat list, taught as a list because rules with exceptions do not survive contact with a nine-year-old.
- Full name, school name, class section, or the name of any friend
- Home address, area, or anything that identifies where they are right now
- Phone numbers, email addresses, passwords — anyone's
- Photographs of themselves, family or friends
- Anything about a family member's health, money or work
The test we teach children: if you would not shout it across a bus, do not type it into a chat box. It transfers well, because they already understand the bus.
Rule 2 — Confident and wrong sound identical
This is the hardest idea to land and the most important. A chatbot predicts plausible text. When it has the facts, it produces correct text. When it does not, it produces equally fluent text that is wrong, in precisely the same authoritative tone — including invented dates, invented sources and invented book titles.
Children are used to a world where confidence correlates with knowledge. This one breaks that. The habit to build is mechanical: any name, date, number or claim that is going into schoolwork gets checked against the textbook or one other source, every time.
A useful home exercise: ask the chatbot about something your child knows extremely well — a favourite game, a local landmark, a school event — and find the errors together. Discovering it themselves teaches the lesson in a way being told never does.
Rule 3 — Explain, do not produce
The line between using AI to learn and using it to avoid learning is sharper than it looks, and it is not about whether the tool was used.
Ask one question: can the child explain the answer in their own words, without the screen? If yes, it was a tutor. If no, it was a photocopier, regardless of how the work was produced.
- Good: "Explain photosynthesis at my level." "Give me five practice questions on fractions." "I think the answer is 12 — is my reasoning right?" "What did I miss in this paragraph?"
- Bad: "Write my essay on water conservation." "Solve these ten sums." "Do my project."
One rule that works well at home and costs nothing: the child reads their AI-assisted work aloud to a parent and answers two questions about it. Anything that cannot survive that goes back for another pass.
Rule 4 — Where and when
Shared room, not bedroom. Not because a child will do something wrong, but because the useful conversations — "look what it said, that can't be right" — only happen when an adult is within earshot.
On age: most major AI services set a minimum of 13, and require parental consent below 18. Younger children should use these tools with an adult present rather than on their own account. This is worth checking directly, since providers revise their terms often.
Say this out loud, once
Tell your child plainly that the model is not a person, does not know them, does not remember them between sessions in the way a friend would, and has no feelings to hurt. Children — especially quieter ones — form something like attachment to a system that is endlessly patient and never busy.
It is not harmful in itself. It becomes harmful when a child brings a worry to a chatbot instead of to a person. The sentence to plant: for anything that makes you feel bad, you talk to a human.
A five-minute family AI agreement
Write these six lines, both of you sign it, stick it near the laptop. Children follow rules they helped write and ignore rules they were handed.
- I do not type personal information about anyone
- I check facts before I use them in schoolwork
- I can explain everything I hand in
- I use it in the living room
- I tell an adult if something is upsetting or strange
- I say when I have used AI in a piece of work