AI Act deepfake disclosure at first exposure: marking is not enough

The Commission's Article 50 FAQ says deployers must disclose deepfake content to a person "upon first exposure at the latest", clearly and in a way people can perceive without special tools, and that deployers "cannot simply rely on the machine-readable marking" the provider embeds under Article 50(2). The viewer has to be able to see or hear the disclosure.

This is a reading of the FAQ, read 2026-10-01, not legal advice. The Sume side is limited to what the docs say about placing a still over a video; see Timeline compose.

What are the provider's duty and the deployer's duty?

The FAQ splits them. Providers of generative systems must make outputs detectable through machine-readable marks (Article 50(2)). Deployers must label deepfakes (Article 50(4)). The two are separate obligations, so meeting the first does not discharge the second.

What counts as perceivable?

The FAQ's examples are visible or audible labels, with no dedicated action needed from the viewer. A metadata field that only a tool can read does not meet that description.

How do I put a visible label on a finished clip?

Sume's compose endpoint takes one still and one video and returns one MP4 with both on screen. In the overlay operation the still is a plate that "keeps aspect", with position (top, center or bottom), width_ratio and margin_ratio controlling where it sits. Both inputs must already be this workspace's media.sume.com artifacts, so import first. Because the plate is composited into the MP4, it travels with the file. Another route is burned text: the captions endpoint accepts authored cues with text, start and end, which skips speech-to-text. The first-frame placement is covered in label burned into the first frame.

What should I check before publishing?

Confirm the label is on screen from the first moment a viewer sees the clip, is large enough to read, and is not covered by other overlays. The Commission's icon page asks for a clearly visible size and no intervening overlay elements. Whether your content is a deepfake at all is a separate question, covered in the three criteria.