EU AI Act Article 50: The Watermarking Deadline on August 2, 2026
7 min read — by HALLMARK.AI
On August 2, 2026, the transparency obligations in Article 50 of the EU AI Act start applying. It is the first broadly applicable law that requires machine-readable marking of AI-generated content — and it is quietly reshaping how the entire content ecosystem treats provenance. Here is what it actually says, who it binds, and what it means if you create or distribute content. (This is general information, not legal advice.)
What Article 50 requires
- Providers of generative AI systems must ensure outputs (audio, image, video, text) are marked in a machine-readable format and detectable as artificially generated or manipulated — the recitals explicitly name watermarking as a technique.
- Deployers of deepfake-generating systems must disclose that content has been artificially generated or manipulated when it depicts real people, places, or events.
- Marking must be effective, interoperable, robust, and reliable as far as technically feasible — the feasibility clause matters, because fragile marks that vanish on the first re-upload arguably don't meet the bar.
Penalties for non-compliance with transparency obligations reach into the tens of millions of euros or a percentage of global turnover, consistent with the Act's general enforcement scheme.
Who is — and isn't — covered
The obligation sits on AI providers and deployers, not on individual creators. If you are a photographer or video creator, Article 50 doesn't require you to do anything. But it affects you anyway, in two ways: the tools you use will start embedding machine-readable marks in AI-assisted output, and the platforms distributing your work are building detection pipelines to honor these labels. Provenance infrastructure is becoming a default assumption of the content stack.
The EU isn't alone
Korea's AI Basic Act — in force since January 2026 — carries a similar labeling requirement for generative AI output, and China's labeling rules for synthetic content run in the same direction. Three of the world's major content markets now mandate machine-readable marking. Whatever your view of the regulations, the direction of travel is set.
What "machine-readable" means in practice
Two technical families satisfy the requirement: metadata standards like C2PA Content Credentials, and invisible watermarking embedded in the pixels. As we cover in C2PA vs invisible watermarking, metadata is stripped by most distribution channels — which is hard to square with Article 50's robustness language on its own. The practical consensus is converging on watermark-plus-metadata: the manifest carries the detail, the watermark survives the pipeline.
The flip side for human creators
A world where AI output must carry a mark is also a world where unmarked-as-AI stops meaning "human-made" — anyone can strip labels, and enforcement is aimed at providers, not thieves. For creators, the durable move is the mirror image of Article 50: embed your own machine-readable mark in your original work, so you can prove authorship no matter what pipeline — human or AI — your content gets dragged through. That is exactly what HALLMARK.AI does: free invisible watermarking for your originals, monitoring for where they end up.