Claude Now Watermarks Its Text, and Copy-Paste Won't Strip It

Claude's text output now carries an invisible, machine-readable watermark, added as part of Anthropic's EU AI Act compliance work. For publishers the practical question is narrow: the marker survives a copy-paste, and it is not a Google ranking signal.

What Anthropic says the marker does

According to a write-up of Anthropic's announcement, the watermark operates at the model level and globally, covering Claude outputs across the API, Claude, Claude Code, Claude Cowork and Claude Tag, and also through the cloud partners AWS, Google Cloud and Microsoft Foundry. The company says it is imperceptible to readers, adds no tokens, and does not change the meaning of the text.

Two further claims matter for anyone handling other people's content. Anthropic says the marker cannot be traced back to an individual, and that it travels with copied text — moving a draft into a CMS, a document or an email does not strip the signal.

The same source says Anthropic's approach aligns with Google DeepMind's SynthID-Text methodology, published in Nature in 2024, and that generated files may carry Content Credentials (C2PA) where supported.

Dates and who can actually detect it

Per Anthropic's help center as cited in the write-up, new Claude models launched in the EU on or after August 2, 2026 mark AI-generated content from day one, and earlier models are being retrofitted. Detection is the part most coverage glosses over: the source says the detection API is in private preview for authorized entities under EU law, which is not general access for publishers, marketers or SEO tools.

Anthropic is not alone in the compliance scramble. OpenAI has published its own position on EU text provenance rules, covering where watermarks apply, how detection works, and why access starts with researchers.

The SEO answer, and the real one

The write-up's SEO conclusion is blunt: a watermark is not itself a ranking penalty, because Google's systems assess whether content is useful and relevant rather than demoting a page for carrying a provenance marker. Its absence equally proves nothing about human authorship.

The operational change is elsewhere. A provenance signal that survives editing shifts the risk from being detected to being unable to say what happened. If an agency delivers copy to a client, or a contractor's draft lands in your CMS, the origin of that text may be establishable later even though nobody can see anything in it today. Teams that already log where AI assisted a draft lose nothing. Teams that treated it as an invisible shortcut now have an auditability problem rather than a quality one.

The practical move is the unglamorous one: write down which tasks may be AI-drafted, require a human to verify factual claims before publication, and keep a record of significant AI assistance where accountability or client communication depends on it. None of that is new advice — the watermark just removes the option of leaving the question open.

Worth noting about the source: the article carries a consultancy pitch, and its claims about the watermark are relayed from Anthropic's own announcement and help center rather than independently tested. Treat the detection-access detail in particular as a statement of policy, not a verified capability.