How we do this

Every review on this site is produced the same way. This page describes that process, including the parts of it that are weak.

Screening checks the channel's most recent 100 uploads for harmful content and videos presenting themselves as something they are not.
A catalogue pass reads up to 120 videos across the channel's full history for subject matter.
A deep review closely examines 10 videos for age-appropriateness and every risk area below.
The AI reads text and still images — it does not watch video or listen to audio, and it can be wrong with confidence.

What we examine, and what we do with it

Channel title and description from the YouTube Data API.
Titles, descriptions, and thumbnail images of the videos we fetch.
What is said in a video, kept only as a capped text excerpt for the review and never the video itself — how and when we can reach it is below.
The uploader's own declarations to YouTube, including whether a video is marked made for children — audited rather than taken at its word.
Every video we fetch is screened for harmful content and for videos presenting themselves as something they are not — the cartoon characters, nursery styling, and familiar titles a young child recognises, wrapped around something they are not expecting. That pattern is commonly called Elsagate, and finding it is why this site exists.
A smaller sample is then read again, much more closely, for age suitability and every risk area below.

How videos are selected, and how many

A channel is read in three passes, and they look at very different numbers of videos. This is the most important thing to understand about a review.
Screening looks at up to 100 of a channel's most recent uploads, one at a time, for harmful content and for videos presenting themselves as something they are not. Every one of those videos is listed on the channel page with what screening found.
The catalogue pass looks at up to 350 uploads and reads up to 120 of them for subject matter, reaching back through the channel rather than stopping at the newest videos. It is run 3 times and only the topics most runs agree on are published, because one run of a model is one opinion and not a measurement.
The review reads 10 videos closely: age band, every risk dimension, and what is said in them. That number is small on purpose — a video read this way costs many times what a screened one does.
So the age guidance and the risk levels on a review describe those 10 videos. The screening flags and the subjects a channel covers describe far more of it. A channel with thousands of videos is still described from a sample, and every review states its own counts under coverage.
Videos our screening has already flagged are always included in the review sample; the rest are spread across the back catalogue rather than taken only from the newest uploads. Full-length videos fill that spread before Shorts do, because a minute of hook and punchline is not what a channel is — but a Shorts-only channel is described by its Shorts rather than not described at all.
The selection is fixed per channel, so re-running a review examines the same videos and the result can be compared with the last one.

What is said in a video, and when we cannot hear it

We can often, but not always, read what is said in a video — as text, not by the model watching or listening to it.
Where we can, we read from across its length rather than only its opening, so a video that turns after the first minute is not judged on the first minute.
Where we cannot, the review says so and the finding rests on the title, description, and thumbnail instead.
A gap is never treated as evidence that nothing was said.
Every finding records what it rests on, so a claim grounded in what was said is distinguishable from one based only on a thumbnail.

What AI can and cannot detect

Analysis is performed by an AI model. ScreenMyYouTuber publishes the conclusions in its own voice and says so on every review.
The model reads text and still images. It does not watch video or listen to audio.
It can miss context, misread humour and tone, and be wrong with confidence.
No review is a safety guarantee, and none replaces an adult watching alongside a child.

What each dimension means

Language: Swearing, insults, and crude language.
Violence: Fighting, injury, weapons, and threat.
Horror: Fear, dread, jump scares, and disturbing imagery.
Mature themes: Sexual, romantic, or otherwise adult subject matter.
Bullying: Mockery, exclusion, humiliation, and cruelty.
Commercial pressure: Selling, undisclosed promotion, gambling mechanics, and scams.
External contact: Pushes viewers to chat, message, or meet somewhere off-platform.
Adult activities: Alcohol, drugs, tobacco, gambling, and driving.
Medical body horror: Injections, surgery, injury detail, and body distortion.
Stimulation: Pace, loudness, and sensory intensity. High is not unsafe, but it is worth knowing before bedtime viewing.
Impersonation: Presents itself as a channel, brand, or creator it is not. A copy of a well-known name is often a different operator entirely, so the reputation a parent is relying on does not apply.
Synthetic media descriptive: How much of this appears machine-generated rather than filmed or drawn by a person. Not a risk on its own, but AI-generated channels can publish at a volume no one is reviewing.
Educational value descriptive: Whether the channel teaches something. Descriptive, not a risk — a high level here is positive.
Consistency descriptive: How alike the sampled videos are. Descriptive, not a risk — high means the sample predicts the rest of the channel well.

What the levels mean

None found: Nothing in the sample raised this dimension. Not a guarantee it is absent from the channel.
Low: Present occasionally or mildly in the sample.
Moderate: Present enough in the sample to be worth knowing about in advance.
High: A prominent feature of the sample.

Confidence is the model's own reported certainty in a finding, not a measure of how accurate it turned out to be.

What the research states mean

Not analyzed yet: No analysis has been run on this channel yet.
Videos screened: Individual videos have been screened, but no channel review has been written yet.
Analyzing: An analysis is running. No review yet.
Full review: An analysis completed and this review reflects it.
Analysis unavailable: The most recent analysis did not complete and no earlier review exists.
Stale: A published review exists, but a newer run is open, failed, or the review is older than the freshness target.
Withdrawn: The review has been withdrawn, or the source channel is gone.

A full review already stands on its own: titles, descriptions, thumbnails, and every declared signal. "Full review + what is said" means this one also rests on what was said in the video, where we were able to reach it — not that reviews without the suffix are incomplete.

Coverage and freshness

Every review states when it was produced.
A review older than 90 days is marked stale.
Channels change. A review describes the sample at the time it ran, not the channel forever.

YouTube's data and our conclusions

Channel and video titles, descriptions, thumbnails, and statistics come from YouTube.
Summaries, age guidance, risk findings, and limitations are ScreenMyYouTuber conclusions drawn from that data.
The subjects listed for a channel say what its videos are about, read from the titles and descriptions of every video we screened. They are not a judgement that the channel teaches those subjects well — that is the educational value finding, which rests on what is actually said in the videos.
ScreenMyYouTuber does not host or play video. Links open YouTube directly.

How to dispute a review

If you represent a channel and believe a review is wrong, contact us and we will re-examine it. A review can be withdrawn from publication immediately on request while that happens.

contact@screenmyyoutuber.org