Three mechanisms get collapsed into one word and they fail differently. A visible watermark is a logo a person can see and a crop can remove. An invisible watermark (SynthID is the named Google version) is a signal in the pixels or samples, built to survive re-encoding. Content credentials (C2PA) are a cryptographically signed manifest bound to the exact bytes of one copy; a platform ingest that re-encodes the file invalidates the hash, which is the mechanism working, not a bug.
Platforms have started to read whatever of this they are given. Meta's 2026 AI labels lean on C2PA/IPTC signals plus self-disclosure. YouTube can auto-label from C2PA metadata when it survives. None of that is the legal disclosure. Article 50 still wants a provider's machine-readable mark on generative systems and a deployer's perceivable label on published deepfakes. Provenance is how you prove a history. The label is how a viewer is told.
For a creator the practical stack is: keep the original export, because that is the copy whose manifest (if any) still hashes; do not expect the uploaded social file to carry it; disclose anyway. A green 'verified' tick in a supporting viewer is a claim about one exact file at signing time, not a property that follows the clip through every re-upload.
In practice
- Archive the generator's export. The social transcode is a different file with a different (or empty) provenance record.
- Use watermark and manifest as complements, not synonyms — one survives upload, one does not.
- A provenance signal is not a substitute for the on-screen AI label or an endorsement disclosure.
The mistake to avoid
Pointing at a C2PA tick after the clip has been through TikTok or YouTube ingest and calling the history intact. The bytes changed; the manifest is gone.
Go deeper
Why C2PA manifests die on upload
A C2PA manifest is a hash over exact bytes, so re-encodes and screenshots kill it by design. How to find the hop in your pipeline where provenance is lost.
Related terms
Content credentials
Content credentials are a cryptographically signed record of who or what made a file and what was done to it, attached by a hash computed over its exact bytes — which is why the record breaks the moment those bytes change, rather than degrading gracefully.
SynthID
SynthID is Google DeepMind's invisible watermark for AI media — a detector-readable signal in the output, not a logo in the corner.
Watermark
Watermark meaning: a mark on generated media that identifies the source, either visible on the picture or embedded in the file.
AI content label
AI content label meaning: a platform or creator tag that marks a post as made or altered with AI. Separate from paid-endorsement disclosure.
Hook rate
Hook rate meaning: the share of people shown a video who are still watching a few seconds in. It grades the opening, not the body.
The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync, in your browser or on your phone.
Free on the web and on iPhone, one account.