EU AI Act Art. 50(2): providers must mark AI outputs; ask your API

Article 50(2) of the EU AI Act requires providers of systems that generate synthetic audio, image, video or text to mark the outputs in a machine-readable format so they are detectable as artificially generated. If you build on an AI video or image API, ask the vendor in writing how its outputs are marked and whether your edits keep the mark.

What the article says

Article 50, read on 2026-10-03, sets the following for providers in paragraph 2. The obligations apply from 2 August 2026, according to the European Commission's code of practice page.

Questions to put to a vendor

The article speaks to providers, but its effects reach customers because your exports carry whatever the provider put in them. Ask:

• Is there a machine-readable mark on image, video and audio outputs, and what format is it?
• Is it metadata in the file, an invisible watermark, or both? Metadata is lost by a re-encode that drops it, while a watermark may survive.
• Which of your edit steps keep it: trim, crop, caption burn, concatenation, resize?
• Is there a documented way to verify the mark on a file you downloaded?
• What changes if you download and re-host the file?

What Sume's docs say

Sume's public docs do not describe a machine-readable mark on outputs, so this post does not claim one. They do document which operations re-encode a file. Video trim defaults to precision: "exact", a frame-accurate re-encode with libx264 and yuv420p, and offers keyframe as a stream copy. Timeline 1.0 renders one MP4 from ordered slots. Any mark that lives in file metadata needs to be checked on the output of these steps, not assumed.

If you need an answer for a specific export route, ask Sume support and test a sample file with a C2PA or watermark reader of your choice.

A testing habit

Pick one generated file and run it through each step of your pipeline in turn. After every step, run the same reader. The first step where the mark disappears is the one to change, or the one to disclose around. A mark that survives your pipeline today can still fail later if a step or a tool version changes, so keep the test as a recurring check, not a one-off.