Behind the scenesJanuary 6, 20269 min read

How we built a million kid-safe YouTube videos with AI

Inside the pipeline we use to discover, transcribe, classify, and human-review over a million YouTube videos for kid safety, at a scale humans alone could never reach.

Filippo YacobBy Filippo Yacob

When parents ask us how Bryte AI can offer a safe slice of YouTube without the rest of YouTube coming with it, the honest answer is a pipeline. A long one. We have built and refined it over years, and it now classifies and curates over a million videos with a level of nuance that would take a team of human reviewers a lifetime to match. AI does the heavy lifting. Humans make the calls that matter.

Step one: discovery

We start by mapping the universe of channels children might plausibly land on. That includes the obvious kids brands, but also long-tail educators, makers, story channels, music channels, and niche hobby channels that show up in search results for things kids actually ask about. Each candidate channel gets a profile that captures topic mix, upload cadence, language, and audience signals.

Step two: transcription at scale

Every video we consider is transcribed in full. Not the auto-captions YouTube ships with, our own transcripts, optimised for kid speech, noisy audio, songs, and code-switching between languages. Transcripts give us something the thumbnail and title never will: what is actually said, minute by minute.

Step three: multimodal classification

Transcript, thumbnail, frames, title, description, and channel context all feed into a classifier stack. We score each video across dozens of dimensions: age band suitability, educational value, narrative tone, presence of advertising or product placement, sensitive topics, scary imagery, unsafe behaviour, and more. No single signal decides anything. The combination does.

Step four: policy routing

Scored videos are routed through our policy engine, which decides whether a video is safe for the youngest band, safe with constraints, safe only inside specific themed pods, or unsuitable. The same video can be approved for one age band and rejected for another. The policy is explicit, versioned, and reviewable.

Step five: humans where it counts

AI is great at processing a million videos. It is not great at edge cases involving culture, sensitivity, or context. Anything the classifiers flag as borderline goes to a human reviewer trained in our standards. So does every flagged report from parents. Humans also spot-audit the AI's confident decisions, because confidence is not the same as correctness.

Step six: ongoing review

A safe channel today can become an unsafe one tomorrow. Creators change direction, take sponsorships, or upload one bad video. We continuously re-score recent uploads, watch for sudden shifts in topic or tone, and pull videos quickly when standards slip. The library is alive, not a one-time approval.

Why AI makes this possible

Without AI, a kid-safe YouTube layer at this scale is simply not buildable. The cost would be prohibitive, the latency between upload and review would be measured in months, and the consistency would drift with every new reviewer. With AI doing the bulk classification and humans doing the judgement calls, we can offer parents something that did not really exist before: a million videos kids love, curated to a standard we can defend, refreshed every day.

That is the answer to the question. Not magic, not luck, just a long pipeline run carefully, with the right jobs given to the right kind of intelligence.