Anthropic has signed the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content. In practice, text produced by Claude will carry an imperceptible watermark, and generated files including .png, .svg, and .jpg will ship with signed C2PA provenance metadata. The company states that marking will apply “wherever Claude is offered, worldwide”.
What Does the EU AI Act Actually Require?
The real trigger here is Article 50, which the Commission confirms took effect on 2 August 2026. It obliges providers of systems that generate synthetic content to mark that content as artificially produced. The enforcement is not just decorative. The fines can run to €15 million or 3% of global annual turnover, whichever is higher.
Why Apply a Brussels Rule to the Whole World?
The reason behind this is pretty simple. Building one marking pipeline and running it globally is cheaper and simpler than maintaining a separate, unmarked product for non-EU users and policing who sits where. So the EU writes the rule and the cheapest corporate response is to make it the global default.
This has happened before. The clearest case is the GDPR, the EU’s 2018 data-protection law. Rather than run two privacy regimes, companies such as Microsoft and Apple rolled out GDPR-style policies to all their users worldwide, and countries from Japan to Brazil built their own laws on the same template.
Does Anthropic’s Watermark Prove Anything?
Unsurprisingly, there has been a lot of media discourse after Anthropic’s announcement, but the results may be less significant than people are suggesting.
Anthropic was candid that a detected mark is a signal, not proof. This means it confirms only that content “may have been processed by Claude”, not that Claude wrote it, since people routinely use the assistant to translate or proofread their own work. The mark can also be stripped by editing, format conversion or a screenshot.
Yet, none of this makes the move cynical. Provenance is becoming a reputational asset, and Anthropic gets to look principled while doing the low-cost route. Google DeepMind already ships SynthID and OpenAI has moved slowly. Complying loudly is good positioning as well as good citizenship.
This is still an achievement for Brussels. Every time a global firm picks one worldwide standard to satisfy the Commission, the EU’s rulebook wins by default, and the lighter-touch American model loses ground which it never actually chose to defend.
Author: Grace Sharp
See Also:
What Is Sovereign AI? Definition, the Money Behind It, and Europe’s Reality
What Did the EU’s Digital Omnibus Actually Delay in the AI Act?
