Check
Claude Watermark Example
Claude watermark example searches end in frustration because there is genuinely nothing to look at.
A marked passage is visually identical to an unmarked one. That is not evasiveness on Anthropic's part — it is the design requirement, stated openly.
There is no visible example. Invisibility is the specification, not an implementation detail.
Runs in your browser. No signup, no upload, nothing stored.
What you would be looking for
In a statistical watermark, the signal is which word appeared at each point where several would have served.
Read one sentence and every choice looks entirely ordinary, because it is. The pattern only exists across hundreds of such choices, and only becomes visible against a key you do not have.
Why an example cannot exist
Anthropic's stated requirement is that the mark does not alter meaning, quality, style or readability.
A watermark you could point to in a paragraph would fail that test immediately, and would be trivially removable by anyone who noticed it. Both properties rule out a visible example, permanently.
What people are usually shown instead
Articles illustrating this topic tend to show a paragraph with highlighted zero-width characters, rendered as little coloured boxes between words.
Those screenshots are real, but they are of a completely different phenomenon. They have left a large number of people confident about a mechanism Claude does not use, which is why the myth has been so durable.
The file case is different
For images there is something concrete to look at: open the C2PA manifest and read the signed claim as structured data.
That is the closest thing to a viewable Claude mark that exists, and it is genuinely inspectable — issuer, claim generator, signature and all.
A useful thought experiment
If you could see the mark, you could delete it. If you could delete it, it would tell nobody anything.
The invisibility that makes this frustrating to look at is the same property that makes it work at all. Any design that satisfied the curiosity would fail the purpose.
What you can demonstrate instead
Generate the same prompt twice and compare the outputs. You will see two plausible passages with different word choices, and no way to tell which selections were nudged.
That exercise conveys the mechanism better than any screenshot could, precisely because it shows you the thing you cannot see.
Source
The claims on this page are drawn from primary documentation and reporting rather than from other tools’ marketing copy.
Anthropic: How Claude marks AI-generated content →Related
Check what your text and files actually carry
Runs in your browser. No signup, no upload, nothing stored.