Claude Now Watermarks Text and Adds C2PA Metadata to Images

Starting August 2, 2026, Anthropic began marking Claude-generated content with two separate signals: an invisible watermark built into the text itself, and C2PA Content Credentials attached to generated image files. The change was reported by TechCrunch and confirmed in Anthropic's own Claude Help Center documentation.

It is easy to read this as one feature. It is really two, with different mechanics, different failure modes, and one of them, the file-level metadata, is the one this site's inspector can actually check.

What did Anthropic announce on August 2, 2026?

Anthropic began applying two marking mechanisms to Claude output: a text watermark embedded at the model level, and C2PA-standard provenance metadata attached to generated image files. Both apply to any Claude model that launches on or after August 2, 2026, and Anthropic says older models are being retrofitted, though it has not given a timeline.

The trigger is the EU AI Act's Article 50 transparency obligations and the Code of Practice Anthropic signed onto, both of which require machine-readable marking of AI-generated content. Anthropic chose not to geofence the feature to European users. The same watermark and metadata ship globally, on every supported Claude surface.

How does the Claude text watermark work?

Anthropic's documentation describes weaving "an imperceptible watermark directly into the text itself." It is not a footer, a disclosure line, or a change to formatting. The signal sits inside the statistical structure of the generated words, so it does not alter meaning, tone, or readability, and it travels with the text wherever it is pasted.

That is the watermark's main advantage over file metadata: copy Claude's output into an email, a document, or another app, and the watermark rides along, because it is part of the text rather than an attribute of the file that carried it. C2PA metadata, by contrast, lives in the container format and does not survive a copy-paste into plain text.

The watermark applies across every Claude surface: the API (Claude Platform), claude.ai, Claude Code, Claude Cowork, and Claude Tag. It also extends to Claude models running inside AWS Bedrock, Google Cloud, and Microsoft Foundry, though Anthropic notes those cloud environments may not carry the separate file-level C2PA metadata, since that piece depends on how each platform handles generated files.

How does C2PA metadata work for Claude-generated files?

For supported file types, currently SVG, PNG, and JPG, Claude attaches signed provenance metadata that follows the C2PA standard, the same manifest format this site was built to read. A C2PA manifest is cryptographically signed metadata bundled into the file itself, recording who or what created it, when, and with which tool, in a way that is tamper-evident rather than merely self-reported.

This puts Claude in the same camp as Google and OpenAI, who committed to shipping C2PA metadata alongside their own watermarking systems in May 2026. It is not a proprietary Anthropic format. Any tool that reads standard C2PA manifests, including this site's inspector, can parse a Claude-generated file's Content Credentials the same way it parses one from Adobe Firefly or a C2PA-enabled camera.

What are the limits of these signals?

Anthropic is direct about what these signals do not prove. Neither the watermark nor the C2PA metadata establishes authorship. A human could have used Claude only to edit or translate existing text, and the watermark would still appear, even though a person wrote the underlying ideas.

The reverse gap matters just as much: a missing watermark does not mean no AI was involved. Heavy editing, converting a file to a different format, or screenshotting an image can all strip both signals. Anthropic itself says it is not yet clear how much editing it takes to break the text watermark specifically, which leaves a real gray zone between "lightly edited, still detectable" and "rewritten enough to erase the mark."

The marking also only covers models released on or after August 2, 2026. Anything generated with an older Claude model, or with a model accessed before that date, was never watermarked or tagged in the first place. There is no way to retroactively add the signal to content that already exists.

How does this compare to SynthID and other AI content marking?

Anthropic is not doing this alone. Google, Meta, Microsoft, and OpenAI have each made comparable commitments tied to the EU AI Act's Code of Practice, and this site has tracked the same pattern show up on YouTube's AI labels, the EU's official AI content icons, and Google Ads disclosure labels.

The closest comparison is Google DeepMind's SynthID, which pairs the same two-part idea, an in-content watermark plus file metadata, but with its own proprietary scheme rather than Claude's. SynthID's text watermark needs Google's own detector to check, and its image and video watermarks are undetectable without Google's tooling. Claude's text watermark is similarly closed for now, but its file-level signal takes the opposite approach: it uses the open C2PA standard instead of a proprietary format, which is exactly why a general-purpose C2PA reader can already parse it.

Frequently Asked Questions