Why YouTube Kids Still Shows Inappropriate Videos

Elsagate and what it revealed about algorithmic filtering

Between late 2017 and early 2018, parents and journalists documented a large-scale phenomenon that became known as Elsagate. Thousands of videos — many reaching tens of millions of views — had used familiar children's characters including Elsa, Spider-Man, and Peppa Pig as thumbnails and titles, while concealing content involving violence, crude humor, body horror, and sexualized scenarios inside. The videos were formatted to look like nursery rhymes and cartoons, and YouTube's automated systems had allowed them to flourish on the Kids platform.

YouTube's response was significant. The company removed millions of videos, hired thousands of additional human reviewers, and tightened the criteria for content to appear on YouTube Kids. By mid-2018 the most egregious examples had been removed and the volume of reports dropped sharply. Many news cycles declared the problem solved.

What Elsagate actually revealed, however, was not a temporary failure that could be patched. It exposed the structural tension at the heart of every algorithmic filtering system: the same engagement signals that make a video perform well for children — familiar characters, bright colors, high click-through rates — are also perfect camouflage for bad actors who understand how to game them. As long as the recommendation engine rewards engagement and the filter works reactively by flagging content after it appears, the gap between upload and review remains open. Content creators who want to exploit that gap have every incentive to keep trying new formats.

The Elsagate playbook has evolved since 2018. Modern versions tend to avoid the most obvious trigger words in titles and use slower, subtler escalation. They are harder to catch in automated review and take longer for human reviewers to identify. The category of "content that looks fine in a thumbnail and first thirty seconds" is essentially unbounded, and there is no automated system in existence that can fully close it.

How the YouTube Kids recommendation engine creates ongoing risk

YouTube Kids is not a static library of pre-approved videos. It is a subset of YouTube's broader catalog, filtered by signals the platform uses to determine child-appropriateness, with a recommendation engine running underneath. That engine watches what each child clicks, how long they watch, and what they watch next, and adjusts what it shows them accordingly. This is exactly how the main YouTube algorithm works — the Kids version simply has a tighter filter on top.

The problem is that "tighter filter on top" and "no recommendation drift" are not the same thing. If a child watches a lot of one particular style of video, the algorithm surfaces more content that resembles it. If that drift takes the child toward content that is borderline — not obviously inappropriate, but not quite right either — the filter may not catch it, and the recommendations continue compounding. Over a session of thirty or forty minutes, the content a child is seeing at the end may bear very little resemblance to what they watched at the start.

YouTube has acknowledged that its automated systems are imperfect. In its own communications about the Kids app, the company notes that it cannot guarantee that all content is appropriate and asks parents to monitor viewing. This is not a criticism of YouTube's engineering team — it is an honest description of what machine learning can and cannot do. Classifying the appropriateness of a video with full context requires human judgment, and there is far too much content uploaded every day for every video to receive human review before it goes live.

New uploads are particularly vulnerable. A channel with a clean history can publish a video that passes automated review and appears in recommendations before any human watches it. In a high-volume content category like children's entertainment, where creators publish daily or multiple times per day, the volume of new uploads alone keeps the review backlog meaningful. Parents who have set up YouTube Kids and moved on are working with a snapshot of the platform, not a stable curated library.

There is also the issue of search. Unless a parent has switched a profile into Approved Content Only mode, children can search for anything within YouTube Kids, and the results are filtered by the same automated signals — not by a parent's prior review. A child who searches for a favorite character may find content that passed automated checks but has not been seen by the parents at all.

Why the allow-list model closes the gap that filtering cannot

The structural limitation of all filtering systems is that they are reactive: content must exist before it can be evaluated, and something must trigger a review. An allow-list inverts this entirely. Nothing can appear in your child's view unless you have explicitly added it. There is no recommendation feed surfacing unreviewed content, no search reaching into the broader YouTube catalog, and no algorithm optimizing for engagement in the background. The universe of content your child can reach is precisely the set of videos and channels you approved.

Watchly is built on the allow-list model. When your child opens the app, they see a library you built: channels you searched for, videos you previewed, and curated libraries you imported by age and topic. When a video ends, the player returns to the library rather than queuing an algorithmic next-up. There is no sidebar suggesting related content and no autoplay chain that can drift into unreviewed territory.

On top of the allow-list, Watchly runs an AI-powered content scan against 21 categories — including profanity, violence, sexual references, frightening content, and drug references — so if a creator you approved later publishes something off-tone, it surfaces for your review before your child can see it. The AI works as a second layer of confidence, not as the primary gatekeeper. The primary gatekeeper is the fact that nothing gets in unless you let it through first.

Setting up a library takes less time than most parents expect. You can import a pre-built library for a specific age group — say, educational science channels for 8-year-olds — with a single click, then add or remove individual channels from the parent dashboard. Once the library is built, the ongoing maintenance is low: you add a new channel when your child asks for one and occasionally glance at the watch history dashboard to see what they have been watching. There is no feed to continuously police and no algorithm to keep resetting.