It is not yet clear whether it's good news, but it definitely forces a rethink of priorities. UNICEF's latest report, based on a ten-country analysis, captures a reality many in the tech industry suspected: children are embracing artificial intelligence at a staggering pace. As many as 20 million minors have already used AI tools, with an adoption rate over three times higher than that of adults.

The UN agency does not mince words, calling the situation a "global experiment" in which an entire generation is growing up immersed in technologies whose full implications are not yet understood, while governance systems remain dramatically behind. This isn't just a sociological data point—it directly affects anyone designing, distributing, or managing AI applications that may be used by under-18s. From virtual assistants to educational platforms, from coding environments for kids to general-purpose chatbots used for homework, the landscape is vast and still largely unregulated.

For organizations building these tools, protecting minors' data is not a legal footnote but an architectural constraint. Regulations such as Europe's GDPR or the U.S. COPPA impose strict requirements on the collection, processing, and storage of data belonging to school-age individuals. When data flows to third-party public clouds, the chain of control lengthens and the risk of non-compliance multiplies. It is no surprise that many companies are evaluating on-premise deployment of LLMs and AI pipelines precisely to limit the exposure perimeter and guarantee auditability.

Self-hosted, on-premise adoption allows keeping data within one's own infrastructure, reducing the attack surface and simplifying compliance demonstrations. Of course, this choice is not free: it demands in-house skills, hardware investments (often GPUs with ample VRAM to handle inference and fine-tuning), and a careful TCO analysis. But for those developing solutions aimed at minors, the trade-off tilts decisively toward control over cloud convenience.

The speed at which children adopt AI also raises transparency and auditing issues. When a language model is used in educational settings, it becomes essential to trace how responses are generated and which data influence them. A local deployment, where the organization has full visibility across the stack, makes it easier to integrate logging, content filtering, and human review mechanisms that would remain opaque in an externally managed environment.

Against this backdrop, the AI-RADAR community is mapping frameworks and best practices for those evaluating the shift to on-premise or hybrid LLM infrastructures, tackling precisely the challenges of sovereignty, compliance, and true cost. Mass adoption by a generation of AI natives is no longer a future scenario but a current fact that raises the bar for anyone designing AI with safety in mind from the very first phase.