How it works, and its limits

Clip Preflight gives you a quick risk estimate before you upload a gameplay clip. It is not YouTube, it can't see YouTube's systems, and it can't promise anything about how your video will be treated.

Your footage stays on your computer

The video is opened and analysed inside your browser tab. It is never uploaded to us or anyone else. The first check downloads two open-source models, an image model and an audio model, plus the code that runs them. That's about 150 MB from public hosts (Hugging Face, Google Cloud Storage and jsDelivr). After that, your browser usually caches them. Those downloads don't include anything about your video. On a metered connection, use Wi-Fi for the first check.

How accurate is it?

Honestly, we don't know yet. This is an early version. So far we've tested it only on synthetic clips we generated ourselves. It hasn't been measured on a labelled set of real gameplay, so expect both misses and false alarms. Always look at the flagged frames and listen to the flagged ranges yourself. We'll publish accuracy figures here once we've measured them.

Check 1: the opening

YouTube's advertiser-friendly content guidelines treat graphic violence more strictly when it appears at the start of a video. We sample a frame every half-second from the first 15 seconds and score each one with an open image model (CLIP) for how much it resembles graphic violence (blood, gore, injuries). Frames above our threshold inside the first 7 seconds (the window YouTube's guidelines name for "no ad earnings") make the verdict red. Frames from 7 to 15 seconds make it yellow, because YouTube's gaming guidance lists that range for limited ads. The same guidelines say blurred or obscured violence is treated more leniently.

Limits: the model looks at single frames, not motion or context. It can miss violence that's small, dark or brief, and it can flag red-heavy scenes (sunsets, neon, red cars) that aren't violent. Always look at the flagged frames yourself.

Check 2: likely music

We run an open audio model (YAMNet, trained on Google's AudioSet) across the whole clip, about once per second, and mark the time ranges that sound like music. Ranges that also sound like singing or rapping are marked as likely vocals, because licensed songs (in-game radio, for example) are a common source of copyright claims. YouTube's own help pages explain that when a copyright claim is made, the claimant may be able to monetize the video.

Limits: this is "likely music detected, may trigger claims". It is not Content ID, and it can't tell whether a specific song is registered with Content ID or who owns it. A game's original score may or may not be claimed. Loud sound effects can sometimes look like music, and quiet background music can be missed.

The verdict

• Red: possible graphic violence inside the opening window.
• Yellow: borderline opening frames, or likely music somewhere in the clip.
• Green: no high-risk signals in what we check. Still not a guarantee: YouTube looks at much more (title, thumbnail, language, the rest of the video).

What we don't check

Profanity, titles, thumbnails, descriptions, sexual content, drugs, or anything after the opening for violence. We may add some of these later.

Not affiliated

Clip Preflight works with clips from any game, including GTA 6. It is not affiliated with, endorsed by or sponsored by Rockstar Games, Take-Two Interactive, YouTube or Google.