A UN agency wrote nine rules for AI systems used by children

The classroom note

A school district evaluating an AI reading tutor for young children has no education-specific AI law to check it against, in most countries. What exists instead, and what a district can still use as a checklist, is a set of children's-rights requirements a UN agency wrote for exactly this gap. UNICEF published version 2.0 of its policy guidance on AI and children in November 2021, setting out nine requirements for what it calls child-centred AI, built on a foundation of protection, provision and participation drawn from the Convention on the Rights of the Child.

What the evidence says

The document is not a study; it is a normative framework the agency asked governments and companies to pilot and adapt, and it names its own limits, acknowledging that the guidance does not claim to have all the answers. Its nine requirements move from protecting children's data and privacy and ensuring safety, through fairness and non-discrimination, transparency and accountability, to preparing children for AI and creating an enabling environment for child-centred systems, each with its own recommendations for governments and businesses. UNICEF has continued revising it: the agency's current guidance page, retrieved 16 September 2026, describes a version 3.0 published in December 2025 that expanded the list to ten requirements and added explicit coverage of AI companions and AI-generated child sexual abuse material, changes the 2021 text could not have anticipated.

The implementation question

The mechanism worth naming is that this guidance carries no enforcement power of its own; a district cannot cite it to a vendor as a legal requirement the way it could cite a statute. What it offers instead is a shared vocabulary for a procurement conversation: asking a vendor to show how its product answers each of the nine, now ten, requirements turns a marketing claim of being designed for kids into a specific, checkable set of questions about data use, transparency, and who is accountable when the system gets something wrong for a child.

What holds and what fails

As a checklist for a procurement conversation, the guidance holds regardless of jurisdiction, because it does not depend on any one country's law existing yet. It fails as a compliance standard: no regulator enforces it, meeting it is self-assessed by the vendor or the district, and the document's own three-version history shows it was already out of date on emerging risks, AI companions and generated abuse material, within four years. Editorially, a vendor's claim to follow UNICEF's AI guidance is worth checking against which version, and which of the nine or ten requirements it can actually document.

• Which of UNICEF's nine, or from December 2025 ten, requirements can this vendor actually document meeting, rather than assert?
• Is the guidance being cited the November 2021 text or the expanded December 2025 revision?
• Does anyone at the district treat this checklist as a starting point for questions, rather than a certificate the vendor already holds?

A voluntary framework cannot make a vendor accountable by itself; it can only tell a district which questions it forgot to ask.