Anthropic said on Aug. 11 it’ll embed invisible watermarks in every Claude output, everywhere, to satisfy a European law. Nine days earlier, on Aug. 2, Article 50 of the EU AI Act had come into force, requiring that generative outputs be “marked in a machine-readable format and detectable as artificially generated,” qualified by “as far as this is technically feasible.” Non-compliance is priced at €15 million or 3% of global annual turnover, whichever is higher.

The scope Anthropic chose is the story. Per the company’s help document, the marks “will apply to output from supported models wherever Claude is offered, worldwide.” That covers the Claude app, the Platform API, Claude Code, Claude Cowork, and Claude Tag, as well as distribution through AWS, Google Cloud, and Microsoft Foundry. Text gets an imperceptible signal woven into token selection; images, .svg files, and other assets get signed C2PA metadata, the same provenance standard Adobe and the BBC have been pushing.

This is the Brussels Effect in miniature. Rather than run a two-tier product, Anthropic is exporting EU compliance to every jurisdiction it operates in, which is generally cheaper than partitioning infrastructure and politically useful in Washington.

The technical caveats are the interesting part. Anthropic concedes the text watermark survives copy-paste and light editing but can be stripped by format conversion, re-saving, or screenshots, and that any detected mark is “not conclusive” evidence Claude produced the content. Researchers have already demonstrated image watermarks can be undone, and open-source C2PA-stripping tools are in circulation. Paraphrasing or routing text through a second model tends to erase the signal entirely.

Two Commission task forces begin work in September to fill in the technical detail the statute leaves blank. Article 50’s feasibility carve-out is doing a lot of quiet work until they report.

Sources

Sources