StoryBeam KidsBlog

Dad · StoryBeam Kids builder

What I Know About Recommendation Algorithms, and Why StoryBeam Kids Doesn't Have One

I build software for a living. That's exactly why "no autoplay, no algorithm, no open search" wasn't a limitation I settled for — it was the first decision I made.

By Jason · Published July 16, 2026

Before I was anyone's dad, I was a software developer. I've spent a lot of career hours in the unglamorous plumbing of apps: the parts that decide what a user sees next. So when people hear that StoryBeam Kids has no "next episode" queue, no personalized recommendations, and no search bar that returns whatever's out there, and they assume that's a feature we haven't gotten around to building yet — I understand why. Every app they've ever used trained them to expect it. I want to explain why I left it out on purpose, and why I'd leave it out even if I had unlimited engineering time.

A recommendation system is not a neutral tool that happens to suggest videos. It's an optimization function, and the thing it's optimized for is almost never "what's good for this particular kid right now." It's watch time, session length, and return visits, because those are the numbers that map to a platform's business. Researchers who've studied these systems closely — including a detailed case study from New America's Open Technology Institute on how YouTube's recommendation engine actually works — describe it as trained through reinforcement learning specifically to broaden what a viewer watches and keep them consuming longer. That's not a leaked secret. It's the stated design goal. I've read enough machine learning literature and built enough features chasing engagement metrics myself to recognize exactly what that objective function is doing: it doesn't know or care that the account behind the screen belongs to a three-year-old. It only knows what keeps the session going.

For an adult, that's a manageable trade-off — you can feel a recommendation feed pulling you somewhere you didn't intend to go, and close the tab. A toddler can't do that. There's no skepticism in the loop, no "wait, why am I watching this" moment. Whatever autoplays next just plays, and she trusts that someone chose it for her. That's the entire design problem in one sentence, and it's why I think autoplay and algorithmic "up next" queues are close to the worst possible interaction pattern to put in front of a young child, even when every individual video in the catalog is perfectly fine.

This isn't a hypothetical. It's a documented history.

The clearest early case is what became known online as "Elsagate." In 2017, parents and reporters — including a widely cited New York Times investigation — documented that videos using recognizable kids'-content branding (Frozen characters, Peppa Pig, Spider-Man) were slipping past YouTube and YouTube Kids' filters and autoplaying into disturbing material aimed at very young viewers. According to YouTube's own public accounting at the time, it terminated more than 270 channels and removed over 150,000 videos in response. The point isn't that YouTube didn't try to fix it — it's that a catalog too large to review by hand, combined with a recommendation system built to chase engagement rather than appropriateness, produced exactly this failure mode, repeatedly, for years.

Two years later, in September 2019, the FTC and the New York Attorney General reached a $170 million settlement with Google and YouTube — $136 million to the FTC, the largest COPPA penalty in the law's history at the time, and $34 million to New York — over allegations that YouTube collected personal data from children on child-directed channels without parental consent, data that fed the same targeting and recommendation systems. That settlement is specifically about privacy law, not content curation, but it's worth naming because it's the clearest regulatory confirmation that a federal agency looked at how YouTube's kids' ecosystem operated and concluded it broke the law.

And this isn't just history. As recently as this spring, Fairplay — a children's advocacy organization — organized an open letter to YouTube's CEO signed by over 135 organizations and 100-plus researchers, including Jonathan Haidt, citing an analysis that found roughly 40% of videos YouTube's algorithm recommended after a popular preschool show contained AI-generated content, and that a substantial share of new users' recommended Shorts were "AI slop." YouTube itself acknowledged the underlying dynamic back in 2019, when it changed its recommendation algorithm specifically to stop amplifying "borderline content," and reported a real, measurable drop in watch time for that category once it did. That admission is itself telling: reducing harm to viewers and increasing watch time were, by YouTube's own account, working against each other.

None of this means every autoplay queue is malicious, or that every engineer building one is indifferent to kids. It means the underlying incentive — hold attention — is structurally misaligned with a young child's interests often enough that researchers, journalists, and regulators have documented it as a pattern across nearly a decade, not a one-off bug.

What we do instead

StoryBeam Kids' catalog is closed. Every show in it was chosen and listened to by a human — usually me — before it was added. There's no open search, because open search means the catalog is effectively the entire internet with a filter on top, and filters are exactly the thing that failed, repeatedly, in the history above. There's no algorithmic "up next," because I'm not willing to hand a three-year-old's attention to an optimization function I understand well enough to know what it actually optimizes for. When an episode ends, the player stops. It doesn't reach for something else on my daughter's behalf.

This costs us something, and I want to be honest about that instead of pretending it's free. A recommendation engine genuinely does surface things a manual review process would miss, and a closed catalog will always be smaller than an open one. StoryBeam Kids will never have as much content as an app that lets an algorithm pick for it. That's the trade we made, deliberately, with our own kid as the test case. My daughter co-hosts the flagship show with her mom because this app started as something we built for our family first. The "no algorithm" rule isn't a marketing angle. It's the same rule I'd want if someone else were building the app she used.

Sources

Jason is a software developer, father to the founder of StoryBeam Kids, and reviews every show in the catalog before it appears.

← What Her Brain Is Actually Doing When She Listens With Her Eyes Closed The "Two-Hour Rule" Isn't a Real Rule. Here's What the AAP and WHO Actually Wrote. →