AI at home
Homework, media literacy and age limits: how families can deal with AI without either banning it or waving it through.
Two kinds of use
| Use | Effect |
|---|---|
| "Write my essay" | The task is gone, and so is the learning |
| "Explain to me why that is so" | A patient explainer, available at any hour |
| "Ask me questions about this topic" | Revision without an argument at the kitchen table |
| "Check my text and say what is unclear" | Feedback that is otherwise missing |
| "Is what I wrote here correct?" | Careful: the answer tends to agree |
The last row is the treacherous one. Models agree too readily. To "is this correct?" a yes comes more often than it is earned.
What children should know
- It knows nothing, it continues text. Which is why wrong answers also sound right.
- It invents sources, numbers and quotations, and convincingly.
- It does not know you and means you neither well nor badly.
- What you type in may be stored by the provider.
- No names, addresses, photos of others and no secrets belonging to other people.
A rule that holds day to day
- 01
Explaining and revising are allowed
The route to the answer, not the answer.
- 02
Handing it in is not
What counts as your own work stays your own work.
- 03
It gets said either way
Whoever used it says what for. Openness rather than control.
- 04
It always gets checked
Every number, every date and every source gets looked up.
Age limits and accounts
| Typical provider rule | Detail |
|---|---|
| Use from 13 | With a guardian's consent |
| Own account from 18 | Below that through a family account |
| Child protection settings | Present, but of varying effect |
Under data protection law the age for valid consent sits between 13 and 16 depending on the member state. In Germany and Austria a guardian's consent is required below the respective age.
What applies at school
The school's rule takes precedence over the rule at home. Where a school requires disclosure, it is disclosed. Where it forbids use, it is not used. See Education.
Why detection tools do not work
Tools that try to detect whether a text came from a model work on statistical signals such as low variance in word choice. Those signals also appear in people writing in a second language or simply writing evenly.
- The error rate for non-native speakers is documented as high.
- A false accusation does harm that cannot be taken back.
- A result is a suspicion, not proof, and carries no sanction.
- Task formats that make the process visible work better than any detection.
Data protection for minors
- Children's data enjoys special protection, and information must be given in language a child understands.
- Profiling for advertising purposes towards children is to be judged especially critically.
- Chat histories can contain health, family and school data. Those are sensitive holdings even when they arise incidentally.
- Withdrawing consent must lead to deletion. Whether a provider implements that is in its privacy notice, not in its marketing.
Accompanying rather than banning
A ban moves use to where nobody is looking any more. What works better is shared use with a clear rule and a regular conversation about what went wrong. The more interesting question at the kitchen table is not "did you use it?" but "where did it talk nonsense at you?".
The family agreement
FREE ACCOUNT
Family agreement to fill in
A single page for the fridge, with the rule, a conversation starter, and what children should know.
Templates2 items