No public checker can confirm a Claude watermark. That is not a gap in the tooling that will be filled next month — it is how the scheme works. Reading the watermark requires Anthropic's key, and that key has not been published. Anthropic's own documentation is also explicit that "nothing is added to the text and there are no hidden characters", so there is nothing in your file for a scanner to find either.
What you can check is a different layer, and it is worth checking: invisible characters that really are bytes in your file, and the C2PA credential Claude attaches to the images and files it produces. Both are visible to anyone with the right reader. Neither is the watermark.
If a site offers you a Claude watermark detector, it is either running an AI-detector guess dressed up as a watermark check, or it is selling you something it cannot deliver.
Claude's watermark is statistical: a bias in how the model picks each next word. There is no marker to find, no byte to scan for, no signature at the end of the file. Detecting it means comparing the text against the pattern the model was configured to produce, and doing that needs the key Anthropic holds.
Anthropic states that detection is available through an API in private preview, offered to organisations that need to verify watermarking — regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups, plus enterprises with their own compliance obligations. It is not a public endpoint and there is no self-serve checker.
Both are real, both are worth doing, and neither is the watermark.
Checking these tells you something true about your file. It does not tell you whether Claude wrote it.
AI detectors — Turnitin, GPTZero, Originality.ai and the rest — are real tools and they do return a score. What they do not do is read a watermark. They look at phrasing tells and statistical properties of the writing, which is a completely different method. Anthropic calls this "fundamentally different from checking for a watermark".
The practical consequence: a detector's verdict is not evidence about watermarking, in either direction. A low score does not mean the text is unwatermarked, and a high score does not mean a watermark was found. And because those tools read the writing itself, cleaning invisible characters out of your file will not move their number.
We cannot tell you whether a piece of text carries Claude's watermark, and neither can any other public tool. Anyone who says otherwise is either misinformed about what a statistical watermark is, or hoping you will not check. The one thing we can do is tell you which layer a given check is really looking at — which is the part most sites get wrong.
Anthropic's position is documented and quoted on this page; where our own measurements appear on other pages, those numbers come from runs on this machine and are not estimates. The two layers explained in full →
No. Reading the watermark requires Anthropic's key, which is not public, and Anthropic's documentation states that no characters are added to the text — so there is nothing in your file for a scanner to find either. What you can run is a check for the two layers that are real bytes: invisible characters, and C2PA metadata on files Claude produced. Both are worth checking, and neither answers the watermark question.
Most of the objection is not about the mark itself but about what it is assumed to lead to: a permanent, invisible record that outlives the document and can be read by someone you did not agree to tell. That is a reasonable thing to worry about, and it is worth separating from the technical question. Anthropic states the watermark carries nothing that identifies you, your organisation, or your chats — but the concern is about who can eventually read it and under what rules, not about what is in it today.
Not by inserting anything. Each time the model chooses its next word it picks from a ranked list of candidates; a statistical watermark nudges that ranking in a direction that is invisible in any single choice but measurable across hundreds of them. That is why it survives copy-paste, translation and light editing — the pattern is in which words were chosen, not in any characters — and also why you cannot find it by looking at the file.
The watermark is not a ranking signal, and Google's public position on AI-assisted content has been about quality rather than about how the text was produced. So on the narrow question, nothing in the announcement changes what is safe to publish. The real risk is elsewhere: if you are publishing at volume without editing, that was already a quality problem before watermarking existed. The longer answers →
That is the speculative end of this, and it deserves a straight answer: nothing in Anthropic's published documentation describes watermarks acting as triggers that other systems react to automatically. What is described is detection for verification — an organisation checking whether content was model-generated. Treat the machine-to-machine version as a hypothetical. If it ever arrives, it will arrive as a published specification, not as something already running.