AI ACT GUIDE
Updated 25 August 2026 · 6 min read
Marking AI-generated content (C2PA, metadata)
Article 50(2) requires machine-readable marking. Transition until 2 December 2026 for systems already online.
Generated images, audio, video and some texts must carry detectable marking, where technically feasible. The code of practice points to C2PA and IPTC/EXIF metadata.
Systems placed on the market before 2 August 2026 get until 2 December 2026 for this technical duty. The user-facing notice is already in force.
The Clairia scanner reads machine-readable marking (C2PA/JUMBF, IPTC/XMP) on a homepage image sample. With no visual classifier, missing markers do not prove an image is not AI-generated.
The path, in order
- 1
Scan the homepage
Paste the URL. In under a minute you see whether a known chatbot is present, and whether a notice is visible. With no chatbot there is no /100 score.
- 2
Show the notice next to the chat
One script line (w.js) sits on Crisp, Intercom, Tidio and known widgets. On Crisp, the sentence is also shown once in the conversation.
- 3
Re-scan the homepage
Clairia should appear in the findings. This is not a certification: it confirms the script is on the page.
- 4
Let the log run
Each page load with the script is dated. It is not proof that the chat was opened.
Frequent questions
- Who must show the notice: the site or the chat vendor?
- The visitor is talking to your site. The vendor (Crisp, Intercom, Tidio…) provides the tool. The site must inform, at the first interaction.
- Is a notice in the terms or legal notice enough?
- No. Too late and too easy to miss during the chat. Article 50 requires clear information at the latest at the first interaction.
- How do you prove the notice was shown?
- A useful file: a register (tool, vendor, use), a dated log of script loads, homepage scans. A /100 score is not proof.
General information, not legal advice.
Official reference: European Commission Article 50 guidelines ↗