“The internet was never built for kids”.
That’s from SuperAwesome’s founding pitch deck, way back in 2013. It’s the problem SuperAwesome was created to solve, and thirteen years later - while a heap remains to be done - we’ve made some progress. Kids ads are no longer stuffed with cookies, age verification is (sorta) a thing and (sorta) works, most big platforms have at least some kind of kid or teen mode, and legislation has expanded the industry’s duty from the narrow commandment “thou shalt not track” to a more expansive requirement to act in the best interests of young people.
Progress, despite all.
But today’s internet is rapidly becoming a very different beast from that of 2013. It’s becoming an internet constructed entirely by, around, with, and for AI. And AI, too, was never built for young people.
Over the last few months I’ve been speaking to technologists, operators and researchers at the frontier of kids’ use of AI.
My broad conclusion. We’ve made a start. But a huge amount remains to be done. And right now, I’m not sure we’re looking in all the right places.
The Iceberg: AI-ware, everywhere
Discussion of kids and teens’ interactions with AI has tended to focus on direct conversations with a chatbot. That’s the most visible part of any individual’s interaction with AI. But it’s not the only part - and it may not be the most important.
Over time pretty much every part of the engagement stack is going to be AI powered. Content creation, discovery, product recommendation, customer service - all of it’s getting re-written for an AI era and kids or teens will interact with AI hundreds of times simply by using everyday sites and products.
And every single one of those interactions will be held to a legal standard developed before an age of AI. A brand or platform’s duty to uphold the best interests of the child* is not lessened when the entity charged with that duty is artificial.
Five problems to solve
I have an awful lot more still to learn about this space. But here are what strike me as five particular problems that remain unsolved. I’m far from clear that this is an exhaustive list and several of the problems interact in complex ways.
The content problem - ensuring AI produces good content for young people. That content is often AI chats - and as such this is the visible part of the iceberg. How do you make sure that an AI doesn’t lie, is helpful but age appropriate when asked about sensitive questions, doesn’t help with cheating**, and is supportive of teens experiencing mental health problems without validating psychosis? Of the last 9 evaluations of frontier chat apps by Common Sense Media’s Youth AI Safety team, two are “high risk” and five are “unacceptable risk”. Hmmm.
The decisioning problem - ensuring AI makes aligned decisions about young people. Every decision, express or implied, that’s made by a computer is going to end up being made by AI. Consider content recommendation as just one example. What content should a user be shown in their favorite app? How does that change based on their age? Their expressed preferences? How should legal regimes, policy guidelines and platform terms of service show up in its reasoning? How might parents chose to set parental controls when they can do so through conversation, not checkboxes? Huge opportunity, huge complexity.
The data problem - providing AI with quality, compliant data about young people. For very good reasons, we restrict data collection at scale about people under the age of 18. But there’s a tension there - in order to produce age appropriate content and decisions, AI will need to have a detailed model of what young people specifically like, watch and do. Definitionally you do not want to use general audience data to recommend content to twelve year olds. Ultimately that can only come from scaled, consented data.
The governance problem - ensuring auditability and control. AI for youth is going to be a domain like legal or medicine - one with huge applications for AI, but low-to-no fault tolerance (arguably youth tech is even more challenging, since answers are less verifiable than domains like law). So just as with those fields, we’ll need AI system architectures that embed humans on the loop, if not in the loop, that support clear records of decision making, and in which compliance can be audited, not just assumed.
The timeliness problem - moving at the speed of youth, with seatbelts. This is partly but not entirely a subset of the data problem. Youth culture moves fast. Even understanding what a teenager is saying may require a detailed understanding of what happened on Discord, Reddit and TikTok, yesterday. Data, policies and judgements will need to combine extremely low error rates with extreme dynamism to remain relevant and effective.
Building Below The Iceberg
This litany of problems may make it sound like I’m pessimistic. I’m not. I can imagine ways in which AI enables a vastly better internet for the next generation, one in which nuanced judgements about the best interests of the child are applied with a sensitivity and intelligence we’ve never been able to achieve at scale before.
But there’s a lot to build, and much of that building will need to happen below the iceberg’s waterline.
SuperAwesome’s going to be busy. Who else is? I'd love to hear from you.
___________________________________________________
* Enshrined, for example, in the California Age-Appropriate Design Code Act.
** There is ofc a whole sub-field of content generation problems to think about in the context of edtech.
AI disclosure: The images in this article are of course, AI generated. Claude read a previous draft of this article, told me it was “seven out of ten” and attempted to re-write it in ways which I didn’t like, and I then re-wrote the whole thing largely from scratch one caffeinated morning. So, excepting the image and a couple of specific legal references, this ended up being all my own work.