Can I see Claude's watermark by looking at the characters?

No. There is nothing to see. The mark is a statistical property of which words were chosen, not a character or a piece of metadata in the file.

Does deleting invisible characters remove the watermark?

No. They are unrelated mechanisms. Removing invisible characters is worth doing on its own terms, but it has no effect on a distributional watermark.

Then what does the checker on this site find?

The things that are genuinely present and genuinely checkable: zero-width characters, non-breaking and exotic spaces, provider-specific HTML class names and data attributes, and typographic punctuation. It runs in your browser and reports exactly what is there.

Is Claude output watermarked right now or not?

Not from any shipped model. The policy applies to models launched on or after 2 August 2026, and Opus 5, Sonnet 5, Fable 5, Opus 4.8 and Haiku 4.5 all predate it. Anthropic describes earlier models as “in progress”, so check the review date on our model list rather than trusting a headline.

Is Claude's watermark invisible characters?

By Claude Watermark Research · Updated August 14, 2026

No. Anthropic's watermark is distributional: a secret key biases which word the model picks among near-equivalent options, so the mark lives in the word choice itself. Invisible characters are a separate, real thing that genuinely does end up in text copied from chat interfaces, and the two are constantly confused because “invisible watermark” in a headline means imperceptible, not hidden characters. The practical difference is large: you can find and remove invisible characters yourself in seconds, and nobody outside Anthropic can currently verify the statistical mark at all.

Check your own text for these artifacts. Runs in your browser, no account, nothing uploaded.

Why the confusion is so common

Headlines this week say things like “Anthropic embeds invisible watermarks into every Claude output”. Read literally that sounds like hidden characters wedged between your words. It is not what the phrase means. “Invisible” is describing that a reader cannot perceive the mark, in the same way you cannot perceive that a coin is very slightly weighted.

It does not help that invisible characters are also real. Zero-width spaces, joiners and byte-order marks do survive copy-paste out of chat interfaces, and people have been finding them in pasted text for years. Both things are true at once, which is exactly why they get merged into one story.

What each one actually is

The statistical watermark is a bias applied during generation. At each step the model has several near-equivalent words available, and a keyed rule nudges the choice. Across a long enough passage that produces a signature. Since the signature is the word choice, it survives copy-paste, retyping and reformatting, and no amount of character stripping touches it.

Invisible characters are ordinary text that happens to render as nothing. They are in the bytes of your document. Any editor, any diff, any search box will reveal them, and deleting them removes them permanently.

So they behave in opposite ways. One is impossible for you to see and impossible for you to verify. The other is trivially findable and trivially removable.

What this means in practice

Claude's marking policy covers models launched on or after 2 August 2026, and Opus 5, Sonnet 5, Fable 5, Opus 4.8 and Haiku 4.5 all launched before that. So the statistical mark is not in text from any model you can use today, whatever the headline says.

Invisible characters, on the other hand, may well be in your text right now, because they come from the copy-paste route rather than from the model. That is the thing worth checking before you submit a document, and it is checkable in seconds.

This also disposes of a whole category of product claim. A tool that strips zero-width characters is doing something real and useful. If it describes that as removing Claude's watermark, it is either confused or hoping you are.

Questions

Can I see Claude's watermark by looking at the characters?: No. There is nothing to see. The mark is a statistical property of which words were chosen, not a character or a piece of metadata in the file.
Does deleting invisible characters remove the watermark?: No. They are unrelated mechanisms. Removing invisible characters is worth doing on its own terms, but it has no effect on a distributional watermark.
Then what does the checker on this site find?: The things that are genuinely present and genuinely checkable: zero-width characters, non-breaking and exotic spaces, provider-specific HTML class names and data attributes, and typographic punctuation. It runs in your browser and reports exactly what is there.
Is Claude output watermarked right now or not?: Not from any shipped model. The policy applies to models launched on or after 2 August 2026, and Opus 5, Sonnet 5, Fable 5, Opus 4.8 and Haiku 4.5 all predate it. Anthropic describes earlier models as “in progress”, so check the review date on our model list rather than trusting a headline.

Check your own text. Free, unlimited, no account, and it runs in your browser so nothing is uploaded.

Open the checker