Law explained

Synthetic content marking: what applies now

Article 50(2): providers of AI systems generating synthetic audio, image, video or text must ensure the output is marked in a machine-readable format and detectable as artificially generated. Techniques named by the law’s recitals include watermarks, metadata, cryptographic provenance (think C2PA-style credentials).

The dates

Machine-readable marking applies since 2 August 2026. Tools already on the market before that date have a transitional period until 2 December 2026 (2026 AI Omnibus). Expect every major AI tool to ship marking by then — many already embed provenance metadata.

Whose job is it?

Primarily the AI provider’s — the maker of your image, video, audio or text tool. Your realistic duties as a business user:

1. Choose tools that mark. Ask, or check the tool’s documentation for C2PA/provenance support.
2. Don’t strip the marks. Exporting, screenshotting, or aggressive re-encoding can destroy metadata. Our Kit’s checklist covers export settings that preserve it.
3. Label visibly where 50(4) applies to you — deepfake-type content and public-interest text are your own duty, on top of the machine-readable layer. Guide →

Where official guidance stands

The main guidance now exists. The Commission’s Code of Practice on Transparency of AI-generated Content was published on 10 June 2026 (adequacy confirmed 8–9 July 2026), and the final Commission Guidelines on Article 50 transparency followed on 20 July 2026. What continues to evolve is the technical layer — interoperable marking standards and provenance tooling (C2PA and friends) are still moving. That’s precisely why our widget and labeling pack are hosted and updatable — and why we publish every adjustment on the changelog. (This paragraph is honest by design: the duties are settled; the marking technology is not.)

Sources are cited in place; official text: Regulation (EU) 2024/1689. This guide is a technical explanation, not legal advice.