California AB 2713: no deleting AI watermarks, and Sume re-encodes
California's 30 September 2026 release lists AB 2713 as barring deletion of digital watermarks. What Sume's re-encoding trim does and does not say about marks.

California's governor listed AB 2713, the "California AI Transparency Act: system provenance data" bill, among the AI bills he signed on 30 September 2026. The release describes it as prohibiting deletion of digital watermarks and allowing access to metadata that shows whether content is AI-generated or user-generated (press release, read 2026-10-02). It gives no effective date, so plan your pipeline now instead of waiting.
What the release says, and what it leaves out
The release is a summary, not the bill text. Read the enrolled text before you make a legal claim; this page only covers what the governor's office published.
Because the release names no effective date, a good habit is to treat provenance as part of the deliverable from today: file, job id, and any marker or metadata that arrived with it.
| Bill | Release summary | Effective date in release |
|---|---|---|
| AB 2713 | Prohibits deletion of digital watermarks; allows access to metadata to tell AI-generated from user-generated content | Not stated |
| SB 1000 | Strengthens the California AI Transparency Act so Californians can identify AI-generated material | Not stated |
| SB 1111 | Digital replicas: likeness protections against false impersonation and deepfakes | Not stated |
Where a re-encode meets a watermark rule
A cut that re-encodes writes a new file. Sume's video trim defaults to precision: exact, a frame-accurate re-encode with libx264 and yuv420p, while precision: keyframe is a stream copy. Timeline compose also renders a new MP4 on the worker media runtime.
Sume's docs do not say that a source watermark or provenance record is carried into those outputs, and they do not say Sume adds one. That is a gap to record, not a promise. A re-encode is also not a deliberate attempt to remove a mark, but you should not assume either way without checking the output file.
In practice that means two things for a team that cuts AI video. First, decide up front which step is the last one allowed to touch the file, so the delivered version is the one you verified. Second, when a platform or client asks whether a mark is present, you can answer from a check on that exact file instead of from memory of the original render.
A defensible workflow
- Keep the untouched original artifact next to every edited cut, with the Sume job id of each step.
- Run the vendor's own check on the file you deliver, not on the original. Google, for example, describes a SynthID Detector portal for uploaded files.
- Do not build a step whose purpose is to strip a mark, and do not accept one from a collaborator.
- If you need a visible label, burn it in with a compose or caption job; it supplements a machine-readable mark and does not replace one.
What Sume does not do
Sume does not verify provenance, does not insert Content Credentials, and does not certify that your file meets AB 2713. It returns a job, an artifact URL on media.sume.com, and the parameters you asked for. Compliance still sits with whoever publishes the file.
Sources
Related posts
More in Use cases
- California SB 1111 digital replicas: avatar and face-swap checklist
California's governor signed SB 1111 on digital replicas on 30 September 2026. A checklist for avatar and face-swap work, and what Sume's docs cover.
- Character turnaround sheet by API: front, profile, back views
Make a turnaround sheet by API: render the front view, then send it as a reference for profile and back. Google's 360-view method, and the Sume request.
- China penalized CapCut and Dreamina over AI labels: what to check
China's CAC penalized CapCut, Maoxiang and Dreamina AI over AI labels in April 2026. What it means if your video tool serves viewers in China.
- China platform AI labels: confirmed, possible and suspected
China asks platforms to sort content as confirmed, possible or suspected AI. What triggers each label and what a video team can do about it.
Written by Sume