If Claude only edited my writing, is it watermarked?

Why editing and generating are not the same case

The mark is applied at the moment a word is selected. At each step the model has several acceptable options and a keyed rule nudges the pick toward one side. Words it never picked were never touched.

When you paste your own draft and ask for corrections, most of the output is a copy of your input. The model is not re-deriving your sentences token by token from scratch; it is returning your words with specific changes. Whatever marking exists is confined to those changes.

This is why “I used Claude to proofread my own essay, is the whole thing flagged now?” has a different answer from “I asked Claude to write my essay”. The two get treated as one question constantly, and they are not one question.

How much signal a light edit actually carries

Anthropic's own stated limits do the work here. Confidence depends on how much marked text there is, which is why short passages carry little and a single sentence carries essentially none. A grammar pass over 1,000 words might change thirty of them, scattered.

Where a correction is forced, there is no freedom to bias at all. Fixing “their” to “there”, closing a bracket, or correcting a name has exactly one right answer, and a watermark needs a choice between near-equivalent options to exist.

So the sparse case is sparse twice over: few changed words, and many of those changes constrained to a single correct form.

Where the line actually sits

Closest to untouched: spelling and grammar fixes, punctuation, formatting, “tighten this without changing my wording”. Small numbers of changed words, many of them forced.

In between: “rephrase this paragraph”, “make this flow better”, “say this more clearly”. Now the model is choosing among genuinely interchangeable options, which is precisely the condition the mark needs.

Fully generated: “write this section”, “expand this into three paragraphs”, “rewrite this in a different voice”. The output is new text and there is no meaningful sense in which it is still yours in the way the first case is.

Most real sessions mix all three, which is the honest answer to where a given document sits: it depends on which parts came back rewritten.

What none of this lets you verify

Nobody outside Anthropic can check any of it. The detection key has not been released, so no tool — including this one — can tell you whether a specific passage carries the mark, and no tool can tell you how much of it does. Anyone selling a percentage is guessing.

There is also a timing question people skip. The policy covers Claude models launched on or after 2 August 2026. Opus 5, Sonnet 5, Fable 5, Mythos 5, Opus 4.8 and Haiku 4.5 predate it, so text from those models is not marked at all, edited or generated. Which model you used matters as much as how you used it, and the list moves — ours was last checked against Anthropic's help centre on 16 August 2026.

What you can check is the other layer entirely: the HTML class names, zero-width characters and exotic spaces that copying out of a chat interface leaves behind. Those are present or absent in the bytes, they attach to anything you pasted regardless of who wrote the words, and the checker on this page reports them exactly.

Questions

Does Claude watermark text it only proofread?

Only the words it actually changed are candidates for marking. Your own sentences come back as your own sentences, and a word-choice watermark has nothing to attach to in text the model did not generate. Anthropic states that detection confidence scales with the amount of marked text, so a light pass leaves very little.

Will my whole document get flagged because I ran it through Claude?

That is not how the mechanism works — it is not a stamp applied to a file, it is a bias in selected words. Text the model did not select carries none of it. Nobody outside Anthropic can verify any specific document either way, so treat confident claims about a whole document as unfounded.

What about asking Claude to rephrase a paragraph?

That is the middle case and it is closer to generation. Rephrasing means choosing among interchangeable wordings, which is exactly the condition a distributional watermark needs. The returned paragraph is model-selected text.

Can I check how much of my document is marked?

No. The detection key is Anthropic's and has not been published. There is no documented way to report a proportion or density of marked content, and any tool offering you that number is inventing it.

Is there anything in an edited document I can check?

Yes, but it is a separate layer. Copying out of the Claude web interface carries HTML class names, and invisible characters can ride along too. Those are deterministic and checkable, and they land in your document whether the words are yours or the model's.