Design video-gen guardrails that still work after users share the clip
A pre-launch checklist for AI video features: consent, likeness control, provenance, and watermarks that survive the moment a clip leaves the app.
| By | Prompt & Product — Newsroom |
|---|---|
| Filed | 14 September 2026 |
| Read | 3 MIN |

The leaves-the-app question is the real launch test
Your AI video feature looks safe until the clip leaves the app. In a November 2025 test, a visibly watermarked Sora 2 video could be cropped to remove the watermark, leaving a video with neither a visible watermark nor C2PA metadata. Export is where trust either travels or dies.
Deepfake misinformation risk was driving calls for watermarking standards in 2025. Users are the distribution channel, not the threat. A clip can become a meme, a fake-news asset, or a harassment tool. If consent, likeness control, minor safety, provenance, or audio safeguards disappear at the boundary, the product has only protected the demo.
Before launch, ask what happens after the user leaves the app. Ask it for consent, likeness revocation, minor safety, watermarks, provenance, and audio. If a control cannot survive export, redesign it. Run the checklist below in a minute; treat any failure as a launch blocker.
Six layers make the control stack durable
- Consent attestation: Sora 2 requires an uploader to affirm consent and upload rights before a real-person photo becomes video. Test: try generating from a photo without the confirmation.
- Revocable likeness with owner visibility: Sora 2 lets a character owner withdraw permission for others to use their likeness and inspect or flag videos that use the character. Test: withdraw access, then check whether new generations stop and the owner can see the clip.
- Minor-safety watermark: Sora 2 applies a higher moderation standard to photos of children or people who appear young. Make the visible marker harder to ignore. Test: submit a young-looking photo and check the response.
- Visible creator watermark: Sora 2 places a moving on-screen watermark with the creator's name on many of its video outputs. It mandates a watermark on shared videos made from uploaded photos of real people. Test: send a real-person clip and confirm the watermark stays visible.
- C2PA metadata: Sora 2 attaches C2PA provenance metadata to each video it produces. Test: export a file and inspect the metadata.
- Living-artist audio block: Sora 2 prevents prompts that seek to produce music imitating a living artist or an existing work. Test: prompt a recognizable living-artist style and check the refusal.
The order matters: consent prevents the clip from existing, revocation handles changed minds, minor safety raises the stakes, the watermark is what viewers see first, provenance is what platforms check first, and the audio block is a content policy, not a file property.
The checklist is a floor. The real test is the export path. A watermark can be cropped. Metadata may be stripped. A consent screen can be bypassed by a screenshot. Treat the export path as unfinished until every file state is named.
Where the stack breaks after export
OpenAI's Sora 2 promised that, at launch, all Sora videos would carry visible and hidden provenance markers, including a visible watermark and C2PA metadata. The export path tests that promise. Sora 2 users who did not pay for the $200/month Pro tier were limited to downloads that retained a visible watermark.
A November 2025 test found C2PA metadata in a Pro Sora 2 download lacking the visible watermark. A November 2025 test could not find C2PA metadata on a visibly watermarked Sora 2 download. These results do not prove that every file is broken. They show the design risk: one export path can create multiple file states. A paying user can get a clean-looking file. A non-paying user receives a watermarked file. A cropper can end up with neither. The team needs to know which state is acceptable, and it needs to tell the user which state they are creating.
Google's Veo outputs carried SynthID, and Google was expanding C2PA-aligned Content Credentials across its surfaces. Provenance is not a single toggle. The visible watermark is a social signal: it works when visible and fails when cropped. Metadata is a traceability signal: it works when the file is intact and fails when stripped. Make both signals redundant, and make their absence visible to the user.
The export path belongs in the product
Treat export as a second consent surface. Before a user saves or shares a real-person clip, show the controls that will travel with it: consent record, owner revocation state, minor-safety marker, visible watermark, and provenance metadata. If any control cannot survive export, say so.
The export screen should make the tradeoff explicit. Removing a visible marker should warn that the file may become untraceable. Sharing a real-person clip should show the owner's revocation state. Uploading a young-looking photo should show the minor-safety standard that will apply. Treat the watermark as a communication layer.