Anthropic confirms Claude now watermarks every output it generates, globally

Anthropic updated its Claude support documentation around August 10, 2026, to confirm a change already rolling out: every piece of text Claude generates now carries an invisible watermark, and every generated image or file, meaning SVG, PNG, and JPG, carries signed provenance metadata built on the C2PA standard. Coverage spans Claude Platform, the API, plus Claude.ai, Claude Code, Claude Cowork, Claude Tag, and the versions of Claude sold through AWS, Google Cloud, and Microsoft Foundry. Anthropic says the labeling applies globally, regardless of where the user is. That is worth knowing now because a large and growing share of production software calls Claude to generate the text and images its own users see.

what Anthropic actually shipped

The text watermark works invisibly: Claude weaves it into the generated text itself, and Anthropic says a reader would never notice it and that it leaves the sense and quality of the writing untouched. It travels with the text through a copy and paste, and can survive some light editing, though heavy rewriting or translating the text into another language will likely strip it out. Generated files, meaning images built as SVG, PNG, or JPG, get different treatment: signed metadata that follows C2PA, the Coalition for Content Provenance and Authenticity standard.

• Claude Platform, the API most production software calls
• Claude.ai, the consumer and business chat product
• Claude Code
• Claude Cowork
• Claude Tag
• The Claude models sold through AWS, Google Cloud, and Microsoft Foundry

the EU deadline behind the global rollout

Anthropic ties that date to its own commitment under the EU AI Act's Article 50(2) Code of Practice, which sets transparency rules for generative AI output. Claude models launched in the EU on or after August 2, 2026 have to carry machine-readable labeling from day one, and Anthropic is retrofitting the same marking onto models that launched earlier. What the Code of Practice does not require is marking output for users outside the EU. Anthropic did that anyway, calling the labeling global rather than EU-only.

Anthropic has not published how the text watermark actually works. Google documents SynthID, its own equivalent for images and audio, in enough technical depth for outside researchers to evaluate it; Anthropic's support article describes the effect but not the mechanism. On the 425-point Hacker News thread discussing the announcement, commenters guessed at methods like token-probability biasing, questioned whether forcing specific tokens into generated code could break syntax or degrade quality, and pointed out that paraphrasing Claude's output, or running it back through a different model, would likely remove the mark.

A detected watermark is not proof Claude wrote the piece from scratch, and Anthropic says so directly. Claude gets used constantly to edit, translate, and summarize text that started out human-written or came from somewhere else entirely, so the mark cannot tell a reader whether Claude generated the whole document or lightly edited someone else's paragraph. The absence of a mark is not proof of the opposite either: text that has been heavily rewritten, translated, or passed through a second tool can lose the watermark without ever having been human-authored in the first place.

Why a build studio cares

GattyWorks ships 24-hour marketing sites and 48-hour full-stack MVPs, built by senior engineers working alongside Claude, including Claude Code for parts of the build itself. Any client-facing copy or generated image that passes through the Claude API now carries this marking by default, whether we call the API directly or a client's own product does. Worth saying plainly before anyone promises a client fully human, undetectable output: for text or images that touched Claude, that promise stopped being true the day this shipped.

Read Anthropic's own explanation of the marking system at support.claude.com, then see where the technical skepticism lands in the 425-point Hacker News discussion underneath it. If your product calls the Claude API and you have not checked whether that changes what you can promise a client about the output, write to us at hello@gattyworks.com.