Two years ago this was a non-issue. Now it's a compliance question, and the answer differs depending on where you post, who's in the clip, and whether you care about the clip staying labelled after someone re-saves it.
Here's the landscape, in the order it actually affects you.
Why this changed
Three things happened at roughly the same time:
- Regulators decided that synthetic media needs to be identifiable, and started writing that into law.
- Platforms started requiring creators to disclose synthetic content, with penalties that range from a label to removal.
- Tooling appeared that can carry a machine-readable "this was AI" signal inside the file itself.
The result is a layered system: a visible mark, an invisible one, and a declaration. Most people only know about the first.
Layer 1: the visible watermark
A logo or text mark burned into the pixels. It survives re-encoding, screenshots, and re-uploads, because it is the image.
Some generators apply one by default; some offer a paid tier without it. Note the distinction that catches people out: a watermark on the free tier is a product decision, and removing it is usually not the same thing as satisfying a disclosure obligation. Turning off a logo doesn't turn off the law.
Layer 2: metadata and content credentials
The invisible layer. Two mechanisms:
C2PA / Content Credentials. An signed manifest attached to the file recording that it was generated, by what, and sometimes with what inputs. It travels with the file and can be read by compatible platforms.
Invisible watermarking. A pattern embedded in the pixels that survives compression and cropping, detectable only with the right key.
The important limitation: both are fragile in practice. Screenshot the video, re-encode it through a messaging app, or record the screen, and the metadata is gone. That's why regulators don't treat metadata as sufficient on its own.
Layer 3: the platform's own disclosure
This is the one with teeth, and it varies by destination:
- YouTube requires creators to disclose when content is realistically altered or synthetic, and applies its own label. Disclosure is in the upload flow, and repeated non-disclosure is treated as a policy violation rather than an oversight.
- TikTok has an AI-generated label, and reads C2PA metadata to apply it automatically where present.
- Meta labels content it detects as AI-generated and asks for disclosure on photorealistic synthetic media.
For a stylised zombie clip, the practical risk is low — nobody is being deceived about reality. But the rules are written for photorealistic deception, and a clip of two real faces doing something they didn't do sits closer to that line than a cartoon would.
What the law says, roughly
Two regimes matter most right now, and both are already in force in some form:
China's labelling rules require both an explicit mark perceivable by users and an implicit mark in the file metadata for AI-generated content, with obligations falling on both the provider and the publisher.
The EU AI Act's transparency provisions require that deepfake content be disclosed as artificially generated or manipulated.
I'm describing the shape of these, not giving legal advice, and the details are still being specified. If you're publishing commercially, read the current text rather than a blog post — including this one.
What to actually do with your clip
For a private clip sent to your partner: nothing. No obligation attaches to a private message between two people who both know what it is.
For a public post: disclose it. One line in the caption — "made with AI from our photos" — satisfies the spirit and, in most jurisdictions, the letter.
For anything commercial: keep the metadata intact, don't crop out a platform-applied label, and check the destination platform's current rule before uploading rather than after.
The part people get wrong
Removing a visible watermark and then posting without disclosure is worse than leaving the watermark on. The watermark is a signal that costs you almost nothing; the disclosure is the actual obligation, and stripping evidence while staying silent is the combination regulators are specifically looking for.
Try it at the generator; price at pricing.
Frequently asked questions
Is a watermark required by law?
Not universally, and not as a logo specifically. What's required in several jurisdictions is that the content be identifiable as AI-generated, by mark or metadata or disclosure.
If the tool removes the watermark, am I clear?
No. A product's watermark policy and your disclosure obligation are separate things.
Does metadata survive re-upload?
Usually not. Screenshots and re-encodes strip it, which is why it isn't treated as sufficient on its own.
Do I need to disclose a private clip?
No. Disclosure rules target publication, not private messages between consenting people.
Does a stylised zombie video count as a deepfake?
It's synthetic media involving real faces. The deception test is what matters in most regimes, and a clearly theatrical clip is far from it — but the disclosure habit is cheap, so keep it.