Anthropic is putting an invisible provenance signal into Claude output, including output from the API and Claude Code.

The immediate reason is regulation. Article 50 of the EU AI Act requires providers of generative AI systems to make generated content detectable in a machine-readable way.

What is remarkable is the scope of Anthropic's response. This is seemingly a worldwide change to Claude. Anthropic's own documentation currently says marking will apply to supported models "wherever Claude is offered, worldwide." It covers Claude, the API, Claude Code, Claude Cowork, cloud providers and generated text generally.

That means a European regulation is seemingly changing the generation path of Claude for an American developer calling the API from Virginia, an Australian company running Claude Code, or a Japanese company using Claude through a cloud provider.

I think this is a dumb move.

There is an uncomfortable alternative too. Suppose Anthropic limited watermarking to the EU instead.

Then Europe would risk putting itself at a technical disadvantage. AI-generated text is not a physical product that has to pass through a European factory. Generation can happen anywhere. Developers can use models hosted outside Europe. They can run open weights on their own hardware. Content can cross a border as plain text.

The AI Act tries to prevent this legally. Its scope can reach providers and deployers outside the EU when their AI output is used inside the Union. That may close part of the loophole on paper. It does not make the technical problem disappear.

A regulator can impose an obligation on a company that does business in Europe. It cannot make every model on Earth emit a watermark.

If compliant European AI services alter their generation while models available elsewhere do not, the regulation has created an incentive to generate elsewhere. If the only practical response is to make every major American model carry the European requirement worldwide, Europe has effectively exported its constraint to everyone who relies on those closed models.

Neither outcome looks particularly clever.

And this is happening while American AI companies are openly worried about losing their frontier advantage to China.

In February, Anthropic accused DeepSeek, Moonshot AI and MiniMax of using more than 16 million Claude exchanges in attempts to distill Claude's capabilities. Anthropic described access to frontier reasoning and coding capabilities as a strategic issue and connected it directly to the American AI lead.

Meanwhile, Chinese labs are shipping models developers can control themselves. Moonshot publishes Kimi's weights. Alibaba has done the same with much of the Qwen family.

Put those developments next to each other.

American frontier labs are fighting to preserve a capability lead over China. At the same time, one of the leading American labs is adding a hidden provenance requirement to its model output, seemingly worldwide, because of European regulation.

Whether that requirement has any measurable intelligence cost is the important technical question.

Anthropic says it does not.

I want to see the evidence.