Anthropic watermarks Claude text: your output is marked

Anthropic confirmed on August 11 that it will watermark text generated by Claude. Not a label in the UI. A statistical mark woven into the tokens themselves, which survives copy-paste and some editing. If your team drafts product copy, support replies or contracts with Claude, that output now carries a signal you did not put there and cannot see.

What actually happened

Per Anthropic's own support documentation, marking works two ways:

• Embedded text watermarks. An imperceptible mark inside generated text. Anthropic says it persists through copy-paste and survives some editing, and does not change readability.
• Signed provenance metadata. For .svg, .png and .jpg output, Claude attaches metadata following the C2PA open standard, which also flags tampering.

Coverage is wide: Claude Platform (API), Claude, Claude Code, Claude Cowork and Claude Tag, worldwide. Models launched on or after August 2, 2026 support marking at launch; older models get it during a transition period. Anthropic says detection documentation is coming so third parties can verify marks.

The driver is regulatory. EU AI Act transparency obligations applied from August 2, 2026, and roughly 190 organisations signed the Commission's Code of Practice on Transparency of AI-generated Content before that date — Anthropic, Google, Meta, Microsoft, Mistral, OpenAI and Synthesia among them. This is not one vendor's product decision. It is the whole model layer moving at once.

Why AI watermarking matters for your business

Read Anthropic's own caveats, because they are the whole story. A detected mark means content may have been processed by Claude. It does not establish authorship. And a missing mark proves nothing — heavy editing, an unsupported format or an older model all produce clean text. So the mark is evidence in one direction only. Anyone planning to police staff or freelancers with it is going to be wrong in both directions, the same trap we flagged when AI text detectors missed styled writing.

Your published pages now carry a provenance signal. If you generate product descriptions, blog copy or category pages with Claude and ship them raw, that text is marked at scale. Nobody has said how platforms will treat it. Spotify decided this week that AI-labelled artists lose algorithmic recommendation. Assume a search engine or marketplace can make the same call, and assume the mark is what tells them.

C2PA on images is the part you should actively want. Signed metadata that detects tampering is useful to you, not just to a regulator. It gives your product photography a chain of custody. The failure mode is your pipeline: most CDNs, image optimisers and CMS uploaders strip metadata on resize. If provenance matters to you, test it end to end before you claim it.

Write it into your vendor and contractor terms now. Two clauses: disclose which model produced deliverables, and do not strip provenance metadata. That costs you nothing today and saves an argument later.

Sources: Anthropic support documentation, European Commission on the transparency Code of Practice, TechCrunch.