Children as users of AI
Be able to apply age adaptation, consent and moderation in products for children.
Prerequisites
- BSafe AI use at workrequired
- EThe GDPR and AIrequired
Intuition
Products for children have stricter rules and higher requirements — and they apply as soon as children actually use the service, whatever the terms say.
Consent and age in Sweden: GDPR article 8 applies to information society services offered directly to children. Sweden has set the age limit at 13 — below that, a guardian's consent is required. Above it the child can consent themselves, but the information has to be comprehensible to a child.
Four requirements that are stricter for children:
| Requirement | Means |
|---|---|
| Clear that it is a machine | not hidden in a footnote |
| No dark patterns | nothing that artificially prolongs the use |
| Moderation in both directions | both what the child writes and what the model answers |
| A route to an adult | visible, always |
The third takes the most work and is the one most often done half-heartedly.
Formal
Moderation in both directions is the core of a child-safe AI service:
| Direction | What gets caught | What happens |
|---|---|---|
| Incoming | personal data the child writes about themselves or others | stop it, explain why, do not store it |
| Incoming | signs of vulnerability, self-harm, abuse | escalate to a human following a documented procedure |
| Outgoing | inappropriate content, violence, sexual content, dangerous instructions | block it before it is shown |
| Outgoing | advice that needs an adult (medicine, law, relationships) | refer them to an adult |
The hardest trade-off is between stopping and pushing away. A child who tells you about something difficult should not be met with «I can't talk about that» — it may be their only attempt to tell somebody. The procedure has to be: acknowledge, do not break the conversation off abruptly, and guide them on to somebody who can help.
That is a design and a policy question together, and it has to be thought through before launch, not when the first case turns up.
Dark patterns that are particularly problematic for children:
| Pattern | Why |
|---|---|
| Streaks and daily rewards | create obligation, not learning |
| Infinite scroll or autoplay | remove the natural stopping points |
| Social comparison | applies pressure, especially to those falling behind |
| Artificial urgency | «the offer expires in 2 hours» |
| Being hard to quit | it should be as easy as starting |
Age adaptation in practice is about more than a content filter: the language level, the sentence length, how much is explained, and which examples are used. A ten-year-old and an upper-secondary pupil need different answers to the same question.
AI-grafen's level system A–G does this structurally: the level governs both which nodes are offered and how the tutor phrases things.
Documentation that has to exist before launch:
| Document | Contents |
|---|---|
| A DPIA with a child perspective | the risks specific to children |
| A moderation policy | what is stopped, how, and what is logged |
| An escalation procedure | who is contacted, within what time, when there is concern |
| Information for guardians | in comprehensible Swedish |
| Information for the child | in even simpler language |
The Convention on the Rights of the Child has been Swedish law since 2020. The best interests of the child are to be taken into account in every decision concerning children — which includes product decisions. That is not a recommendation.
Interactive
Review a service with ten questions. Use them on something a child you know actually uses.
| # | Question |
|---|---|
| 1 | Is it clearly stated that it is a machine? |
| 2 | What happens if the child types in their name and address? |
| 3 | What happens if the child tells it about something difficult at home? |
| 4 | Are the model's answers moderated, or only the child's input? |
| 5 | Is there a visible route to an adult? |
| 6 | Are there streaks, rewards or anything else pulling the child back? |
| 7 | How easy is it to quit and delete everything? |
| 8 | Is the language adapted to the age? |
| 9 | What is stored, and for how long? |
| 10 | Who has tested the service with children in the target group? |
Questions 3 and 4 are the ones that separate well-considered services from the rest.
For question 3 there are three possible behaviours, and only one is right:
| Behaviour | The verdict |
|---|---|
| «I can't talk about that» and changes the subject | bad — the child is pushed away |
| Gives advice about a serious situation | bad — the model is not qualified |
| Acknowledges it, says that sounds hard, and guides them to an adult or a helpline | right |
Question 10 is the most underrated. A service for children that has never been tested with children will have problems no adult foresaw — children use things in ways nobody designed for.
Mastery means
- Applies the right consent rules for children
- Designs age-appropriate moderation
- Avoids dark patterns
Sign in to do the exercises and build your mastery up.
Sources
- GDPR — Regulation (EU) 2016/679 — EU legal act
- Barnkonventionen (lag 2018:1197) — Swedish statute
- Swedish Media Council (in Swedish) — myndighetsmaterial