Do you have to label AI video? The YouTube, TikTok and Meta rules, side by side

What happened

Nothing shipped this week. That's the point.

The AI disclosure rules on YouTube, TikTok and Meta have been in force long enough that most creators have stopped reading them — and started publishing synthetic video at a volume nobody was hitting when the policies were written. The exposure grew. The rules didn't move.

So here they are side by side.

YouTube requires disclosure when AI makes a real person appear to say or do something they did not, when it alters footage of a real event or place, or when it generates a realistic scene that never happened. YouTube also flags AI-generated music as requiring disclosure when that music is the main focus of the video.

TikTok requires labeling for AI-generated or significantly edited content showing realistic people or scenes. Unlabeled content may be removed or have its distribution restricted.

Meta covers photorealistic video and realistic-sounding audio that was digitally created or altered — including realistic AI voices and AI-generated songs with vocals. The rule applies across Facebook, Instagram and Threads.

Three wordings. One shared trigger. If the output looks real, it needs a label — the question is never whether you used AI, it's whether a viewer could mistake the result for a recording of something that happened.

What this is NOT

It is not a rule against using AI. Script help, title generation, captions, thumbnails, color grading and ordinary editing generally fall outside YouTube's disclosure requirement. You do not label a video because an LLM helped you outline it.

It is not a cartoon rule. Clearly stylized or animated content is generally outside the main disclosure requirement on these platforms. The threshold is realism, not technique.

It is not an exemption for fiction. A sketch is still a realistic depiction if it renders a real-looking person in a real-looking place. "Obviously a bit" is your judgment, not the classifier's.

It is not a confirmed penalty schedule. We could not confirm a published, fixed consequence — a set demonetization window, a strike count — for failing to disclose on any of the three. What's documented is that unlabeled realistic synthetic content can be removed or restricted. Treat that as the risk, not a number you can plan around.

And it is not evidence that enforcement went up. We found no confirmation that any of the three changed enforcement volume in the last three days. If someone tells you there's a crackdown this week, ask them for the source.

The real operational risk

It isn't using AI. It's publishing realistic synthetic media without the label at scale.

One mislabeled video is a correction. A short-form operation pushing fifteen clips a day through a template that never sets the flag is a systemic problem — and the cleanup lands on distribution and monetization, not just the single upload.

The second trap: assuming one platform's toggle satisfies another's policy. It doesn't. Cross-posting tools do not reliably carry a disclosure flag from YouTube to TikTok to Reels. Each upload is its own compliance event.

What to do about it

1. Audit the last 30 days of uploads. Flag anything with a synthetic person, voice, location or event that reads as real.
2. Write a one-line test into your workflow: could a viewer believe this is footage? If yes, label it. Stop debating intent.
3. Set the disclosure on each platform at upload, not in a shared template. Verify it stuck on the live post.
4. Treat AI voice and AI music as covered. Both YouTube and Meta name audio explicitly. Voice clones of real people are the highest-risk category you can publish.
5. Give freelancers and clipper teams the rule in writing. Most unlabeled posts come from someone who was never told.
6. Check each platform's current policy page before you build a standing process. These rules have been revised before and the penalty language is the part most likely to change.
7. Keep a record of what you labeled and when. If distribution drops, you want to know whether the label was the variable or something else was.