I read the UNICEF AI guidance. It's good. It's not enough.

Sixty-seven per cent of UK teens now use AI — almost double what it was two years ago. Thirty-nine per cent of American elementary school children learn through AI applications. Those are UNICEF's own numbers, from the front of its third and most serious guidance document yet on AI and children.

This is the third version of UNICEF's guidance on AI and children. It published the first one years ago, worried about algorithms recommending which video a child watches next. By this version, published December 2025, the concern has changed shape entirely — because now the AI doesn't just recommend the video. It makes the video. It writes the essay. It does the homework. And the line that actually stops you cold is the one describing how these systems now claim to be the "friend" of children and present themselves as such.

I want to give this document real credit before saying anything critical. It's built around three plain ideas: protection, meaning do no harm; provision, meaning do good, when it's actually appropriate to use AI at all; and participation, meaning children get real agency in shaping the systems built around them. That's not a bad three-word summary of what a practitioner would want from any system touching a child's life.

Here's my actual problem, and it's not the content — it's the format. This is guidance. Not law. The document itself, in its very first recommendation, calls on governments to build "regulatory frameworks, oversight and compliance" — which is a polite way of admitting that right now, in most places, none of that exists yet.

I've spent twenty years watching what happens to good guidance with no teeth. It gets cited in a footnote of a policy meeting. It doesn't stop the product from reaching a child by Tuesday. This document is thoughtful, well-researched, genuinely trying to take children seriously. And it is, by its own admission, a request — not yet a rule.

The guidance also went further than most policy papers usually bother going: it consulted children directly, across twelve countries, reaching twelve thousand children and their caregivers. That's a genuinely good step, and worth naming. But the document says of itself, plainly, that "children remain at the margin of shaping AI systems" — and that meaningful participation "implies more than inclusion; it demands influence." Being consulted after the fact and being present at the design table are not the same right, even when policy papers describe them with the same word.

Lucie Straková

20-year child welfare practitioner, transitional foster carer, and founder of Best Interest — a channel and essay series on child rights, systems, and AI. More about Lucie →