The mark went in before the reader did. Since August 2, 2026, text generated by new Claude models has carried a machine-readable watermark that Anthropic states does not change meaning, quality or readability. It is invisible by design, to whoever wrote with it and to the client the draft ships to. Anthropic confirmed on August 11, 2026 that it is live worldwide: Claude.ai, the Claude Platform API, Claude Code, Claude Cowork, Claude Tag, and Claude reached through AWS, Google Cloud and Microsoft Foundry. The trigger is European. Article 50(2) of the EU AI Act carries a Code of Practice on Transparency of AI-Generated Content, and Anthropic has signed it. The coverage is not European: it went global nine days later. Generated image files take a different route, with PNG, JPG and SVG carrying signed C2PA provenance metadata instead of a text watermark. Then the sentence that decides what all of it is worth, and it is Anthropic's own: the mark is designed to persist through copying, pasting and "some editing," while the public detection tool is planned and not yet published. The marking has started. The reading has not.
Credibility: HIGH. Three independent trade outlets (TechCrunch, Search Engine Journal, Search Engine Land) all trace back to one triggering artifact, an update to Anthropic's own help-center documentation. This is a company disclosing its own compliance work, not a contested third-party accusation, so the bias risk runs low. The limit worth naming: with no detection tooling in public hands, nobody outside Anthropic can yet test how well the mark survives a real edit.
Fact bullets:
- Effective August 2, 2026: the watermark applies to Claude models released in the EU on or after that date, under the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, which Anthropic signed.
- Global as of August 11, 2026: live across Claude.ai, the Claude Platform API, Claude Code, Claude Cowork, Claude Tag, and Claude accessed via AWS, Google Cloud, and Microsoft Foundry, and not restricted to EU users.
- Text and images are handled differently: an invisible, machine-readable text watermark that Anthropic states does not change meaning, quality, or readability; signed C2PA provenance metadata for generated SVG, PNG and JPG files.
- Anthropic's own stated limit: a detected watermark shows only that content "may have been processed by Claude." It does not prove Claude generated the original ideas, and no watermark does not mean content wasn't AI-generated. A public detection tool is planned but not yet published.
Stay ahead of AI search
The changes that move your rankings and AI citations — and the exact move to make — before they cost you. Free, three times a week.
Until August 11, 2026, an agency's exposure on undisclosed AI drafting was reputational. It is now also technically checkable, and the checking runs on Anthropic's schedule rather than yours.
Read the limit before anyone panics, because it is Anthropic's own: a detected mark shows content "may have been processed by Claude," not that Claude produced the thinking, and an absent mark proves nothing either way. So the useful reading is narrow. A detection result does not expose a page as machine-written; it puts one specific page in question, in front of a client who was already half-wondering. That is a smaller problem than the panic version, and a more awkward one to have. It lands hardest on clients operating in the EU, where the AI Act's disclosure obligations are already in force, but Anthropic applies the mark worldwide, so nobody gets to file this under European compliance and move on.
Recommended actions:
1. Pull the last quarter's Claude-drafted client content and flag the pieces that shipped with light editing rather than a genuine rewrite, starting with any client operating in the EU.
2. Change the content-production workflow so Claude output is substantively restructured before publishing rather than copy-pasted-and-polished, since Anthropic states the mark "may persist through some editing."
3. Record the decision rationale in the production log wherever a client knowingly ships lightly-edited AI drafting. The call is situational per client and jurisdiction, and the log becomes the defensible record once Anthropic's detection tool is public.
TechCrunch: Anthropic says it will watermark text generated by its AI models, TechCrunch, 2026-08-11 — https://techcrunch.com/2026/08/11/anthropic-says-it-will-watermark-text-generated-by-its-ai-models/
Search Engine Journal: Anthropic To Mark Claude Text & Files Under EU AI Act Code, Search Engine Journal, 2026-08-11 — https://www.searchenginejournal.com/anthropic-claude-watermarks-eu-ai-act-code/585355/
Search Engine Land: Anthropic adds AI text watermarking to Claude models worldwide, Search Engine Land, 2026-08-11 — https://searchengineland.com/claude-ai-text-watermarking-484667