Watermarks for AI Text

What Did Anthropic Announce?

On August 2, 2026, the EU's Transparency Code of Practice (part of the AI Act) came into force. It requires providers of AI systems to label AI-generated or AI-edited content so that other systems can automatically recognize it. Anthropic signed the corresponding code of practice in July 2026 and has since been equipping Claude with a technical watermark for text.

This affects every way Claude generates text: the Claude app, the Claude API (Claude Platform), Claude Code, Claude Cowork, and Claude Tag (the Slack integration). Models released from August 2, 2026 onward include the watermark from the start; older models are being retrofitted with it gradually. Images aren't covered by the same process but instead carry metadata via the open C2PA standard.

How Does the Watermark Work Technically?

The method is called SynthID-Text and originally comes from Google DeepMind (published in 2024 in the journal Nature); the theoretical foundation traces back to a 2022 proposal by computer scientist Scott Aaronson. The trick: while writing, Claude repeatedly chooses between several equally valid phrasings. Normally, this choice is random. In the watermarking process, that randomness is instead steered by a cryptographic key that, together with the preceding text, creates a pattern – invisible to humans, but detectable later with the right key.

Detecting it requires three things: the cryptographic key (which only Anthropic holds), a sufficiently long passage of text, and the unmodified original text. The watermark doesn't reveal which user or organization generated the text – only that a Claude model was involved. Anthropic has yet to release a public verification interface for third parties.

What Does This Mean for Your Website Text?

For plain usage, nothing changes for now: the watermark is invisible, doesn't affect the speed or quality of the text, and can't simply be removed by copying the text unchanged. If you value editorially reworked text over text copied 1:1, you can actually tell from the watermark itself:

AI text copied verbatim retains the watermark fully – detectable as soon as a verification tool is available.
Lightly edited text (fixing individual phrasings, copyediting) usually retains the pattern partially.
Significantly rewritten or freshly worded text loses the watermark almost entirely, since the original word choices are no longer present.
Body text with many facts, figures or code is only weakly marked to begin with, since there's little room for alternative phrasing.

Important for context: a watermark isn't a value judgment. It merely reveals how much an AI model contributed to the original text – just as a spellchecker or translation tool is also just a tool in the editorial process. Anyone who fact-checks AI text drafts, enriches them with their own knowledge, and takes editorial responsibility for them is still publishing genuine, credible content of their own.