Claude's text carries a watermark that only Anthropic can read

My take: A mark you can't check proves nothing to you. Here, the proof is in plain sight: Claude drafts, I review, and my merge is the green light.
What to know
- On August 14, 2026, Anthropic explained how it marks Claude's text, with a version of SynthID-Text, a Google DeepMind method: word choices are nudged into a statistical pattern, with no hidden character, so copy and paste keeps it. Light edits probably leave it; a complete rewrite removes it. It applies worldwide; models launched before August 2, 2026 get it "over the coming months", and the page names no model.
- Only Anthropic can read it, through a detection API in private preview (regulators, media, researchers, among others). On November 16, 2026, no public tool checks a Claude text.
- The reason is the EU AI Act. Its Article 50, applicable since August 2, 2026, asks AI providers to mark what their systems generate. Whoever publishes AI-written text to inform the public on matters of public interest must say so, unless a human reviewed it and someone holds editorial responsibility.
- Google's SynthID Detector, open to everyone since October 7, 2026, looks for the watermarks of Google and its partners (OpenAI, NVIDIA, Kakao, Apple) in images, video and audio, not text. A negative result only means Google's AI did not make the file; another AI may have.
Spotted on @ninon.ia_officiel and @it_connect.fr.
What changes
As a reader, see the watermark as a clue in a few hands, not a test you can run: a site that claims to spot Claude's text is not reading this mark, and SynthID's "not detected" never means "written by a human".
As a publisher, the honest label takes one line. Here, Claude drafts, I review, and my merge is the green light (how I write); each post drafted that way ends with "Drafted with Claude, Anthropic's AI, from my take. Reviewed and owned by me." Publish unreviewed AI text that informs the public, and the AI Act asks you to say so.
Source: Anthropic, “How Claude's text watermark works”, August 14, 2026
Drafted with Claude, Anthropic's AI, from my take. Reviewed and owned by me.