EU AI Act Article 50 AI video labeling requirements for creators
Article 50 applies from 2 August 2026: deployers must tell people about deepfakes and providers must add machine-readable marks. What that leaves to a creator.
Article 50 of the AI Act applies from 2 August 2026. The European Commission's guidance page splits it in two: providers of AI systems must add machine-readable marks so AI-generated content can be detected, and deployers must inform people when they are exposed to deepfakes. A creator who publishes generated video is on the deployer side of that split, in my reading; this is not legal advice.
The Article 50 wording below is from the Commission's guidelines page, read 2026-09-29. The Sume facts come from its video docs.
Who has to do what under Article 50?
The Commission page lists the duties by role. It also names the Code of Practice on Transparency of AI-generated Content as a way to comply.
| Who | What the page says |
|---|---|
| Provider of the AI system | Add machine-readable marks so AI-generated or manipulated content can be detected |
| Deployer of the AI system | Inform individuals when they are exposed to deepfakes |
| Provider that follows the Code of Practice | Satisfies the marking and labeling obligations |
| Provider that declines the Code | Must show compliance through alternative equivalently adequate means |
Am I a provider or a deployer if I use a video API?
The page does not sort individual creators into a role, so treat this as a question for a lawyer. The plain reading is that the company that builds and offers the generator is the provider, and the person who uses it and shows the result to an audience is the one informing viewers. A creator who only calls a generation API is not the party that adds the machine-readable mark.
Does a video made with Sume carry a machine-readable mark?
Sume's docs pages on video generation and clip inspection do not say that outputs carry a watermark, C2PA record or other mark. Video inspect's probe reports duration, size, rotation, frame rate, codecs, audio and HDR, not provenance. So do not assume a mark exists: check the file with your own reader and disclose anyway.
What should a creator do before publishing?
Treat disclosure as a step in your own upload, since it is your duty as the party showing the video.
- Decide whether the clip is realistic enough to be mistaken for real footage, a real person or a real event.
- Tell viewers in the place the platform gives you. YouTube, for example, shows its label in the player for photorealistic content and in the expanded description otherwise (YouTube Help).
- Keep a note of which model made each clip, so you can answer a question later.
- Re-read the Commission page before you rely on this post; it was updated on 6 August 2026 and may change.
Sources
Related posts
More in Use cases
- How to extend an AI video past 30 seconds
Sume has no extend parameter on video models. Chain clips: pull a last frame, use it as the next first_frame, then join with Timeline. Vendor limits too.
- Facebook Reels ad specs: 9:16, 1440x2560, H.264 and safe zones
Meta's ads guide for Facebook Reels lists 9:16, 1440x2560, MP4/MOV, 4 GB, H.264 and safe zones of 14% top and 35% bottom. What Sume can and cannot hit.
- Facebook Reels API video length: 3 to 90 seconds, 24 to 60 fps
Meta's Reels publishing docs list 3–90 seconds, 9x16, 1080x1920 recommended, 24–60 fps, closed GOP and 48 kHz AAC. How Sume durations line up.
- Financial advisor video marketing with AI, compliance first
Financial advisor video marketing with AI: short explainers from an avatar, scripts approved before rendering, exact on-screen text, and per-second costs.
Written by Sume