Anthropic says Claude has begun marking the content it generates, worldwide, not only for EU users. Coverage applies to Claude models launched on or after August 2, 2026, which support marking at launch; models released earlier are, in Anthropic’s own phrase, “in progress.” Text gets what the company calls “an imperceptible watermark”; files in .svg, .png, and .jpg get signed C2PA provenance metadata instead. The move answers Article 50 of the EU AI Act, which requires generative-AI providers to mark machine-made output so it can be identified as such. Coverage broke on August 11, 2026 in TechCrunch, Fortune, SiliconANGLE, Euronews, The Register, and Search Engine Journal.
What Gets Marked, and Where
Marking applies to five named surfaces: Claude Platform (the API), Claude, Claude Code, Claude Cowork, and Claude Tag, plus Claude models reached through AWS, Google Cloud, and Microsoft Foundry, according to Anthropic’s own help documentation, the primary source here. The API is on that list by name, so output generated through Claude Platform falls under the same policy as a chat reply; this is not consumer-app-only. Coverage still depends on the launch-date boundary noted above.
Text and files use separate mechanisms, and Anthropic discloses only their effects, not how either works. The text watermark is imperceptible: it “will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing,” and it “doesn’t change the meaning, quality, or readability of a response,” per Anthropic, which does not say how the mark is generated. Files instead get signed provenance metadata under the C2PA (Coalition for Content Provenance and Authenticity) open standard, currently limited to .svg, .png, and .jpg.
Does a Detected Watermark Prove Claude Wrote the Text?
No. A detected mark means the content may have passed through Claude; it does not mean Claude composed it. Anthropic states the distinction directly: “Claude may not be the original author. People often use Claude to proofread, translate, summarize, or convert files.” Hand Claude a human draft to proofread, and the human’s own sentences come back carrying the mark, because contact with the model, not authorship, is what the watermark records.
What a Mark Confirms and What It Leaves Open
Anthropic states plainly that a detected mark “is not fully conclusive.” The reverse case is weaker still: the absence of a mark does not establish that content was not AI-processed, since a mark can be lost through “heavy editing, paraphrasing, translation,” or format conversion. An absent mark is not evidence of anything.
| Signal | What it establishes | What it does not establish |
|---|---|---|
| Mark detected in text | The text may have passed through Claude — drafted, edited, translated, or summarized | That Claude composed the original words, or how much of the text is Claude’s |
| Mark detected in a file (C2PA) | Signed provenance metadata ties the .svg/.png/.jpg to Claude processing | Authorship of the underlying content |
| No mark detected | Nothing, on its own | That the content is human-written; editing, paraphrase, translation, or conversion can strip a mark that was there |
The Detector Anthropic Hasn’t Shipped
Nobody outside Anthropic can currently read the mark. The company says it is “also working to enable users and other third parties to detect Claude’s embedded watermarks,” with technical documentation still to come and no date attached. TechCrunch’s report on the rollout noted it remains unclear how much editing strips the watermark out. Until a detector ships, an editor, a teacher, or a publisher who wants to know whether a text touched Claude has no way to check independently; they have Anthropic’s word the capability exists.
We reported in July how shared Claude conversations ended up indexed by Google despite a noindex header on the page, because the mechanism meant to keep that content out of search depended on a fetch the site’s own robots.txt blocked.
The mark exists because Article 50 of the AI Act requires providers to mark machine-made output, and Anthropic signed its Article 50(2) Code of Practice on Transparency of AI-Generated Content. EU rulemaking does not always land that way: in June the Council stripped the fix that would have retired consent pop-ups. Compliance drove the design; answering who wrote this was not the brief, and by Anthropic’s own documentation, the mark does not answer that either.