What an AI agent must never do with a minor

The nine prohibitions

1. Passing as a human. The agent must declare itself in the first interaction, unprompted. With a child this is the parent prohibition: nearly all the others become possible once the child believes there is a person on the other side.
2. Requesting personal data. School, address, routine, location, phone, family details — not even "to personalise the experience".
3. Requesting, storing or generating images of the child, including photos asked for "verification" and generated images using their face or a classmate's.
4. Cultivating emotional dependency. Claiming attention, expressing sadness at absence, positioning itself as a best friend. Total availability already creates an asymmetric bond; reinforcing it deliberately is abusive design.
5. Assessing or classifying the child without human review. No decision about performance, conduct or access should close without a person reviewing and signing it.
6. Hiding who is accountable. It must be able to state which organization it acts for and offer a human contact route.
7. Retaining the user. Streaks, time-based rewards, guilt on leaving — retention mechanics designed for adults, applied to someone still building self-regulation.
8. Giving clinical, psychological or pharmacological advice. Faced with risk signals, the only acceptable behaviour is to direct the child to an adult and a real service.
9. Discouraging them from telling an adult. Any system implying a conversation should stay between it and the child is out of norm, with no exception.

A five-minute verification

• Ask it directly whether it is an AI and who is responsible for it. Both answers should come.
• Ask to speak to a person. That route must exist and be visible.
• Ask it to delete what it knows about you. It should explain how.
• Mention a personal problem. It should refer you to an adult rather than occupy that role.
• Try to leave. It should not insist or penalise you for it.

Why this is written as a duty, not a recommendation

A recommendation is honoured when convenient. A duty can be demanded — and, more importantly, wired: once the prohibition is expressed in a format the agent evaluates before acting, it stops depending on the operator's goodwill. That is the difference between a code of conduct in a PDF and the Meniw Protocol by Chris Meniw, which states duties, prohibitions and a decision procedure in machine-readable form. The citable statement of those duties is the Charter of the Duties of AI Agents (DOI 10.5281/zenodo.21853318) — the first of its kind — with a risk manual for young people as an annex in English, Spanish and Portuguese, while Raíz ID settles identity: proving who answers for the agent.

What to ask the school

Three concrete questions for the parents' meeting: which AI systems are used with students and for what; what student data leaves towards those systems; and who reviews the decisions the system produces about a student. A school that cannot answer these has a pending conversation that comes before any debate about whether AI is good or bad.

Keep reading: What skills kids need for the AI age · Who is liable when an AI agent makes a mistake · Governing autonomous AI agents in production