EU AI Act Article 50: marking AI-generated content
If your app generates images, video or audio that people in the EU see, Article 50(2) asks you to mark every output in a machine-readable way: from 2 August 2026, or by 2 December 2026 if your system was already on the market. One Kennmarq.com call adds both layers the Code of Practice names.
In short
- Who: providers of generative AI systems, wherever they're based, when the output is used in the EU (Art. 2(1)(c)).
- When: since 2 August 2026; systems already on the market by 2 December 2026.
- What: machine-readable marks. The Code of Practice asks for at least two techniques (signed metadata and a watermark), a free detector, and terms that forbid removing the marks.
- Kennmarq.com: both marks and the free checker in one call; a terms clause to copy is below.
- Fines: up to EUR 15 million or 3% of worldwide turnover (Art. 99).
Who and when
- Providers of generative AI systems, wherever they are based: the Act covers providers outside the EU when their system's output is used in the EU (Art. 2(1)(c)).
- Since 2 August 2026 for systems placed on the market from that day. Systems already on the market before it must mark by 2 December 2026 (Regulation (EU) 2026/1744).
- Fines for breaching Article 50 reach EUR 15 million or 3% of worldwide annual turnover, whichever is higher; for small and medium companies, whichever is lower (Art. 99).
What the Code of Practice asks, and what we add
The Code of Practice on transparency of AI-generated content was finalised on 10 June 2026. Signing it is voluntary and gives a presumption that you meet Article 50.
| The Code asks | Kennmarq.com |
|---|---|
| At least two machine-readable techniques, such as signed metadata and an imperceptible watermark | A C2PA manifest, signed and time-stamped, and an invisible watermark: TrustMark in pictures, Video Seal in clips, AudioSeal in songs |
| A free way for anyone to detect the marks | The checker for pictures, in the browser, and POST /v1/verify |
| Terms that forbid users to remove the marks | Yours: a clause to adapt is below |
| Evidence that the marks hold up | Our robustness report |
Every picture and clip we return carries both layers, read back from the finished file before it leaves; if either fails, the call fails and the mark isn't counted. Songs always carry the credential, and the watermark whenever the sound can hold it.
One call
curl https://api.kennmarq.com/v1/mark \
-H "Authorization: Bearer $KENNMARQ_API_KEY" \
-F file=@render.png \
-F source=created \
-F provider="Acme Inc." \
-F app_name="Acme Studio" \
-F model="FLUX.1 schnell" \
-o render.marked.pngAdd -F markets=eu to keep a signed receipt of what each file carries for the Act, and what stays with you.
A terms clause to adapt
The Code asks providers to forbid users to strip the marks. A starting point, to adapt with your counsel:
You may not remove, alter or obscure the watermarks, Content Credentials or other markings that identify outputs of the service as generated or altered by artificial intelligence.
Deepfakes: a visible label (Art. 50(4))
Deployers who publish deepfakes must disclose them so people can see it; the Code makes the capitalised "AI" the mandatory label. Our marks are machine-readable only, so the visible label is yours for now; a visible label option is planned.
What stays with you
- Being the provider of your AI system.
- Declaring the right source for each file: made by AI, or edited with it.
- Telling people when they talk to an AI system (Art. 50(1)), and labelling deepfakes (Art. 50(4)).