If children are moving from videos to AI tools, what does child safety by design look like?

I keep coming back to one finding in Ofcoms Childrens online experiences research report: children are still encountering harmful content through personalised feeds. That matters because the child who opens an app looking for a homework helper, a game, or an AI tool is often entering the same recommendation system as the child who opened it for entertainment.

The shape of the internet is changing quickly. Kasperskys latest report on childrens digital interests says AI tools moved from sixth to third place in children's Google searches, and online learning entered the top five. So the question is no longer only how to keep children away from the worst content. It is what a child-facing digital environment should look like when curiosity, learning, chat, and entertainment all live in the same product stack.

Ofcom's research covers what changed before and after the new children's safety duties came into force in July 2025. The part that stands out to me is how often the route to harm is structural. The report says personalised feeds remain the most common route to harmful content, with comments and other surface features also playing a role. Skys summary of Ofcoms findings makes the same point in plainer language, saying the feeds themselves still need more change.

That is why age checks and parental controls matter, even though they sit a level below the real design problem. The LSE and 5Rights report argues that more of the industrys recent movement has shifted toward end-user tools, while protective defaults have become less common. It also says age assurance is now widespread, but implementation remains inconsistent and largely unaudited. That feels like a familiar pattern in consumer software, responsibility gets pushed outward unless the architecture is built differently from the start.

From a builder's point of view, that is the interesting question. If children are using AI tools for search, curiosity, schoolwork, and play, then child-facing AI should have different defaults from general consumer AI. That might mean narrower discovery, tighter sharing, clearer stopping points, stricter limits on persistence, and much stronger adult oversight. The exact version will depend on the product, but the principle is straightforward: the experience has to be designed for children, not adapted after the fact.

That is close to how we think about oodlü. Our focus is on architecture first, with adult-created groups, restricted communication, and boundaries built into the world itself. We are interested in systems where the structure does some of the work, rather than asking families or moderation teams to carry everything after launch.

We do not have the policy answer, and we do not have the classroom answer. We are looking at this as a software team trying to understand what changes when children's digital discovery starts to shift from videos toward AI tools. If you build for children, regulate platforms, or work with families in practice, what would you put in the must be designed in from day one column?

We’d love to hear your thoughts on this. Find us on the social channels linked at the top of the page.