Article 50: creator duties vs provider duties, side by side
Article 50(4) puts deepfake and certain text disclosure on deployers, while 50(2) puts marking on providers. A two-column split for teams that publish AI video.

Under Article 50(4), a deployer that publishes a deepfake must disclose that it is artificially generated or manipulated, with lighter disclosure for evidently artistic or creative works. AI-generated text on matters of public interest needs disclosure unless a human reviewed it editorially. Article 50(2) is the separate provider duty to mark outputs in machine-readable form. If you publish video, you will usually sit in the first column.
The split
This is a reading aid built from the article text on the linked page. For an actual decision on your content, read the article in full and the Commission's guidance.
| Question | Provider, Art. 50(2) | Deployer, Art. 50(4) |
|---|---|---|
| Core duty | Mark outputs machine-readably so they are detectable as artificially generated | Disclose deepfakes as artificially generated or manipulated |
| Quality bar | Effective, interoperable, robust and reliable, as far as technically feasible | Disclosure to the audience |
| Creative works | Not the focus of this paragraph | Limited disclosure |
| Text | Marking applies to synthetic outputs | Public-interest AI text needs disclosure unless human editorial review |
| Exception | Assistive editing that does not substantially alter input | Editorial review exception for text |
Where a generation API sits
A service like the Sume video API produces video from a prompt. You, the caller, decide what to publish and where. That puts the publication-side decisions on you: whether a clip is a deepfake, whether the creative-work wording applies, and how the disclosure appears.
What counts as a deepfake depends on the content. A realistic generated person who looks like a real one is a different case from a stylised mascot. Treat that call as a documented decision, not a default.
A deployer's working file
Keep one row per published clip. It takes five minutes per clip and gives you something to show if asked.
- Show the disclosure at first exposure, not buried at the end.
- Write it in plain words: "AI-generated" or "synthetic".
- For text, record who reviewed it and when.
- Revisit the file whenever the code of practice or guidance changes.
clip,job_id,realistic_person,creative_work,disclosure_text,placed_at,date
launch-9x16.mp4,<job id>,yes,no,AI-generated video,first frame caption,2026-10-03What this post leaves out
It does not cover penalties, member-state enforcement or transitional rules. Those are in the regulation and the Commission's pages, and in other posts on this blog.
Making the call on creative work
The creative-works wording is the part teams most want to use and most often misread. It limits how the disclosure is shown; it does not remove it. A film-like piece can carry a short, unobtrusive notice rather than a banner over the picture.
A practical approach is to write two disclosure forms for each campaign: a plain one for ads and a lighter one for creative pieces. Choose per clip, and record the choice.
Revisit the choice if a clip moves to a new context. A scene that is clearly art on a gallery page may read as real news footage when cropped and shared elsewhere.
Sources
Related posts
More in Use cases
- Case study video narrated by an avatar: three scenes, 55 seconds
Narrate a written case study as a 55-second, three-scene Sume avatar video: problem, change, result. Cost by tier and a body you can send; no names needed.
- Cyber Week five-day creative sprint: a daily Seedance clip budget
Shopify defines BFCM as Thanksgiving through Cyber Monday. Twenty-two 8-second 720p vertical clips across those five days cost $101.6928 on Sume.
- Cybersecurity Awareness Month 2026: four avatar clips for $11.04
CISA's 2026 theme is Securing the Next 250. Four 15-second Sume avatar clips, one per behavior, cost $11.04 on Standard or $14.70 on Plus.
- Demand Gen carousels: 2 to 10 matching cards from one reference
Demand Gen carousels take 2 to 10 cards. Image assets run 4:5 or 9:16 at 5 MB. How to batch matching cards from one reference image with the Sume image API.
Written by Sume