Does Claude's watermark survive copying and pasting?

Yes. Anthropic says the Claude text watermark is part of the text rather than separate metadata, so it travels with the words when they are copied and pasted elsewhere, and may persist through some editing. Heavy rewriting or translation can still wipe it out, so the mark is a signal rather than a guarantee.

Which Claude models carry the marking?

Claude models launched on or after August 2, 2026 support marking at launch, according to Anthropic's help document. Anthropic adds that it is also working on models it has already released, so older Claude models should gain output marking during the transition period the EU allows. Coverage is therefore partial today and widening.

Can anyone detect Claude's watermark today?

Not yet, but Anthropic has now said a watermark detection API is coming and that it is working out the details of the implementation. Until that ships, no public tool reads Claude's text watermark. Anthropic also cautions that detection works poorly on short passages, because a few words leave the detector too few of Claude's word choices to judge.

Does the watermark make Claude's answers worse?

Anthropic says the Claude watermark does not change the meaning, quality or readability of a response. In the SynthID-Text paper that introduced the technique, Google DeepMind found no statistically significant differences from the unwatermarked model, and human raters comparing watermarked and unwatermarked answers side by side saw no difference in quality.

Does Claude's marking apply outside the EU?

Yes. The EU AI Act is the trigger, but Anthropic states marking applies to output from supported models wherever Claude is offered, worldwide. It also covers Claude served through third-party providers such as AWS, Google Cloud and Microsoft Foundry, so a US or Asian developer calling the Claude API gets the same marked output as an EU one.

Which file types get C2PA metadata from Claude?

Anthropic lists .svg, .png and .jpg as the supported file types for signed provenance metadata, which follows the C2PA open standard. Plain text output is covered by the watermark instead. A valid signed C2PA record indicates that Claude processed the file and shows whether the provenance data was tampered with afterwards.

Anthropic · 2026-08-10 · major

Claude marks its output — Anthropic adds text watermarks and C2PA file metadata

Anthropic now marks Claude output. Supported models weave an invisible watermark into generated text, and generated .svg, .png and .jpg files carry signed C2PA provenance metadata. Marking covers the API and every Claude product.

Anthropic help-center illustration for how Claude marks AI-generated content

Claude output now carries a hidden text watermark and signed C2PA metadata on generated files.

Quick facts

Maker	Anthropic
Text marking	SynthID-Text watermark inside the wording
File marking	Signed C2PA metadata on .svg, .png, .jpg
Covered models	Launched on or after August 2, 2026
Where it applies	API, Claude, Claude Code, Cowork, Claude Tag
Third-party hosts	AWS, Google Cloud, Microsoft Foundry
Detection	Detection API announced, not yet live

What is it?

Anthropic has started marking what Claude generates. Supported Claude models embed an imperceptible watermark inside the text itself, so the mark travels with the words when someone copies and pastes them somewhere else. Generated .svg, .png and .jpg files get signed provenance metadata instead. Models launched on or after August 2, 2026 support marking at launch, and Anthropic says it is adding marking to models it already released.

How does it work?

The text watermark is a version of SynthID-Text, the approach Google DeepMind published in a Nature paper in 2024. Picking the next word normally draws on a random number; watermarking replaces that draw with a value worked out from a secret key plus the few words already written. The wording stays varied and the meaning does not change, but the run of choices follows a pattern a detector can check afterwards, and a longer passage gives the detector more choices to read. For files, Claude attaches signed provenance metadata that follows the C2PA open standard, so a valid signed record can show that Claude processed the file and whether someone altered that record.

Why does it matter?

EU AI Act Article 50 requires providers of generative systems to mark output in a machine-readable format that is detectable as artificially generated, and this marking is Anthropic's answer to it. The marks apply worldwide rather than only in Europe, and reach third-party hosts including AWS, Google Cloud and Microsoft Foundry, so anyone building on Claude ships marked output by default. Anthropic warns that a detected mark is not conclusive evidence and that a missing mark does not prove AI was not involved.

Who is it for?

developers shipping Claude output

Frequently asked questions

Does Claude's watermark survive copying and pasting?: Yes. Anthropic says the Claude text watermark is part of the text rather than separate metadata, so it travels with the words when they are copied and pasted elsewhere, and may persist through some editing. Heavy rewriting or translation can still wipe it out, so the mark is a signal rather than a guarantee.
Which Claude models carry the marking?: Claude models launched on or after August 2, 2026 support marking at launch, according to Anthropic's help document. Anthropic adds that it is also working on models it has already released, so older Claude models should gain output marking during the transition period the EU allows. Coverage is therefore partial today and widening.
Can anyone detect Claude's watermark today?: Not yet, but Anthropic has now said a watermark detection API is coming and that it is working out the details of the implementation. Until that ships, no public tool reads Claude's text watermark. Anthropic also cautions that detection works poorly on short passages, because a few words leave the detector too few of Claude's word choices to judge.
Does the watermark make Claude's answers worse?: Anthropic says the Claude watermark does not change the meaning, quality or readability of a response. In the SynthID-Text paper that introduced the technique, Google DeepMind found no statistically significant differences from the unwatermarked model, and human raters comparing watermarked and unwatermarked answers side by side saw no difference in quality.
Does Claude's marking apply outside the EU?: Yes. The EU AI Act is the trigger, but Anthropic states marking applies to output from supported models wherever Claude is offered, worldwide. It also covers Claude served through third-party providers such as AWS, Google Cloud and Microsoft Foundry, so a US or Asian developer calling the Claude API gets the same marked output as an EU one.
Which file types get C2PA metadata from Claude?: Anthropic lists .svg, .png and .jpg as the supported file types for signed provenance metadata, which follows the C2PA open standard. Plain text output is covered by the watermark instead. A valid signed C2PA record indicates that Claude processed the file and shows whether the provenance data was tampered with afterwards.