The EU Transparency Code changed a frontier model in nine days

The EU AI Act‘s Transparency Code took effect on 2 August, and within nine days it had changed a frontier model. That’s a faster result than most technology regulation manages in a year.

The requirement itself is short: AI companies must mark generated or edited content so other systems can identify it. Not label it for humans, which is a lower bar, but make it machine-readable.

Anthropic responded by watermarking every Claude model released after that date, embedding the signal into token selection rather than into file metadata. TechCrunch tied the change directly to the Code.

OpenAI, Google, Meta and Microsoft have committed to comparable practices under the same guidelines, and analysis of the C2PA side of it notes files get signed metadata while text gets the embedded mark. So the Code is doing what regulation is supposed to do and rarely does, which is move an entire industry at once.

A European rule changed what American models do worldwide, because nobody builds a separate model for one continent.

How far the EU Transparency Code reaches

That extraterritorial reach is the part US-based builders keep underestimating. Anthropic’s watermark isn’t applied only to European users, because segmenting model behaviour by geography is expensive and fragile.

Which means the practical compliance boundary for a frontier lab is now the strictest jurisdiction it serves. That’s the Brussels effect, and it’s operating on AI exactly as it did on privacy.

Compare it with the American approach in the same fortnight. Four labs met the White House on 4 August to agree a voluntary testing framework, and voluntary is the operative word.

One regime asked and got commitments. The other required and got shipped code within nine days. That contrast is the whole argument about which regulatory model works, running as a live experiment.

What the EU Transparency Code leaves out

The Code’s weakness is what it doesn’t specify. It requires marking, not accuracy about what the mark means, and it says nothing about how detection results should be used by schools or employers.

So a compliant watermark can be technically correct and still get someone accused of something they didn’t do. TechTimes put the distinction well: the mark proves processing, not authorship.

Builders reading this should assume marking obligations apply to anything they ship into Europe, and that the practical answer is to inherit the upstream provider’s mark rather than build their own. Forbes has a longer treatment of what the shift signifies.

Watch the enforcement question next. A code with no test case is a code nobody has priced yet, and the first penalty is what tells the industry how seriously to take the rest of the Act.