Stop Asking Whether. Start Asking How.

Let me be direct about something first.

Unsupervised AI use by children carries real risks. A June 2026 UNICEF brief found that one in ten children surveyed used AI for advice about things that worried them. That is not a homework helper. That is something more relational, and the safeguarding implications of a child forming a confiding relationship with an AI system that has no adult oversight are serious. UNICEF's companion report on AI chatbots describes children relying on AI for information, creativity, advice, and sometimes relationships, and argues for clearer accountability and oversight. That argument is correct.

The Common Sense Media risk assessment of Google Search published in July 2026 adds another layer. It gives AI Overview and AI Mode an Unacceptable Risk rating and notes that these features cannot be turned off by parents, administrators, or users. AI has moved from being a separate tool a child chooses to open, to being the default layer inside the tools they already use every day. That changes the nature of the problem considerably.

So the risks are real, embedded, and in some cases not within a parent's control to remove. That needs saying plainly.

With that said, I want to ask a different question.

The debate that has already been settled

Whether children should use AI is no longer a live question in the way it once seemed. The UNICEF data shows they already are, at scale, across a wide range of uses. This is the same pattern of adoption that we saw with computers, with the internet, with smartphones, with social media. The technology arrived in children's lives before the guidance did.

I believe that a more useful question, one that feels genuinely worth working through, is what kinds of AI use actually help children develop, and what kinds quietly undermine the thinking work that learning requires.

Because those two things are not the same, and the difference matters enormously.

What the research suggests about good use

Wayne and I have spent a lot of time working through the pedagogical research on how children learn best. Some of it has direct bearing on this question.

Productive failure is a learning design developed by Manu Kapur that shows something counterintuitive: children learn more durably when they struggle with a problem before being given the answer. The struggle itself prepares the mind to receive and retain the instruction that follows. Asking AI for the answer first inverts this entirely. The child skips the productive part.

Elaborative interrogation points in a similar direction. Children learn better when they ask themselves, or others ask them, why something is true rather than simply receiving the fact. The act of generating an explanation, even an imperfect one, creates stronger and more transferable understanding than passive receipt of information.

Stop and Think research, including work by Professor Kaśka Porayska-Pomsta and Professor Wayne Holmes whose work informs oodlü's design, shows that requiring children to pause before answering promotes deeper cognitive processing, reduces guessing, and leads to stronger long-term retention.

What these approaches share is a commitment to keeping the thinking work with the child. The child retrieves, struggles, explains, and questions. The tool, whether a teacher, a game, or an AI, responds to that effort rather than replacing it.

AI used passively, as an answer machine, displaces all of this. The child receives without processing. The retrieval does not happen. The struggle, which is where the learning is, gets skipped. The feedback arrives before the thinking has occurred.

AI used actively is a different matter. Asking AI to generate questions rather than answers. Using it to challenge a position the child has already formed. Asking it to argue the other side of something the child has written. These are uses that could support rather than replace the cognitive work. They require the child to have done thinking first.

The homework question

I want to raise this carefully, because it is a question rather than a recommendation, and teachers are far better placed than I am to judge the classroom context.

Are there situations in which AI-assisted work is not just acceptable but genuinely useful to a child's development? Not because it saves effort, but because learning to use AI well is itself a transferable skill that will matter throughout their working lives.

A structured, discrete module on using AI purposefully, understanding its limitations, interrogating its outputs, and identifying where it is wrong, seems different in kind from a general permission to hand over thinking. The goal in that context would shift from whether the child produced the right answer to whether they engaged usefully with the process. That is a meaningful distinction.

Using AI well is a skill. It requires knowing when to trust it, when to push back, how to verify what it says, and how to use it as a starting point rather than an endpoint. Children who develop that skill will have a real advantage. Children who simply learn to ask AI for answers and copy them will have learned very little.

The question worth asking is whether we help children develop that skill deliberately, or leave them to absorb whatever the defaults provide.  I sense that the former is more useful.

None of this resolves neatly. The risks are real, the benefits are real, and the children are already in the middle of it. What feels clear is that the question of whether AI belongs in their lives has been answered by the children themselves. The question that remains is whether the adults around them, parents, teachers, and the people building the tools, help them develop a relationship with it that serves their thinking, or leave them to absorb whatever the defaults provide.

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