Guides
Claude Watermark False Positives When Proofreading
Claude watermark false positive proofreading is the sharpest problem with this system, and it is not a bug in the implementation.
The mark records that text passed through Claude. It cannot record how much of the text Claude actually contributed.
Marking tracks processing, not authorship. Proofread your own writing and it comes back marked.
Runs in your browser. No signup, no upload, nothing stored.
The mechanics of the problem
Ask Claude to fix your grammar and it regenerates the passage, applying the watermark as it produces each token.
What comes back is substantially your writing, and it carries the mark. Nothing distinguishes it from text Claude drafted from a blank page — the signal records the generation step, not the origin of the ideas or the sentences.
Anthropic acknowledges this
The documentation states plainly that a detected mark does not confirm Claude is the original author, and that a mark is not fully conclusive.
The company is not overclaiming here. The risk is entirely in how a future detector's output gets read by people who did not read that far.
Who this hits hardest
Reporting after the announcement identified consistent groups, and none of them are trying to deceive anyone:
- Non-native English speakers who use AI to polish their own prose
- Writers with dyslexia or other accessibility needs
- Students who write their own work and check it before submitting
- Professionals under editorial or client no-AI policies
- Developers whose commit messages and docs are lightly assisted
How to keep authorship provable
Keep drafts. Version history, timestamped documents and commit logs demonstrate authorship far better than the absence of a mark ever could.
If a policy applies to you, disclose the assistance rather than obscure it. Disclosed proofreading is a conversation; discovered concealment is an incident, and the difference is mostly in the sequencing.
The narrow workaround
Ask Claude to list the errors rather than return corrected text. Then apply the fixes yourself, in your own document.
The advice is yours to act on and no Claude-generated prose ever reaches your file. It is slower, and it is the only genuinely clean solution available.
Where this is heading
Once detection ships, the burden will quietly invert: people will be asked to explain a mark rather than to justify suspicion.
That is a meaningful change in who has to prove what, and it lands on exactly the groups listed above. It is worth deciding your own disclosure position before someone else decides it for you.
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.