ClaudeRewrite vs Manual Editing: Which Actually Removes the Watermark?

A structural rewrite removes Claude's watermark; manual editing usually does not. The reason is structural, not about effort: Claude's watermark is a statistical pattern seeded token by token, so unless you change a large fraction of the words, hand-editing leaves most of the pattern in place.

Both approaches keep your meaning intact, so the real question is which one actually breaks the mark. Here is how they compare and why the answer comes down to how the watermark is built.

Why manual editing struggles

Claude's watermark, based on SynthID-Text, biases the model's choice among plausible next words. Crucially, the bias at each position is seeded by the words immediately before it. Change one word and you shift the seed for the next, but you do not undo the overall pattern unless you change enough words across the whole passage.

That is why swapping a few synonyms or varying a sentence here and there is unreliable: a human editor cannot see the statistical bias, so they cannot target it, and light edits leave most of it intact. To remove it by hand you would essentially have to rewrite the document, which is slower than letting a model do it. Anthropic itself states that a complete rewrite is what removes the mark.

Why a structural rewrite works

ClaudeRewrite regenerates the wording through DeepSeek V4 Flash, a different, open-weight model. Because it uses a different tokenizer and sampling strategy and applies no watermark of its own, it rebuilds the token pattern from scratch without re-stamping a new one. It changes structure, order, and phrasing while preserving every fact, number, name, quote, and code block, and an automatic meaning check confirms nothing drifted.

Two rewriting mistakes to avoid: do not rewrite Claude with Claude, because the same watermark can be re-applied, and do not rewrite with Google's Gemini, because Gemini watermarks its own output with SynthID and would swap Claude's mark for Google's.

What about invisible characters and file metadata?

These are separate from the watermark. Text copied from a chat interface can carry stray invisible characters and smart quotes, and Claude-generated images carry C2PA provenance metadata. A manual editor rarely thinks to check either. ClaudeRewrite strips invisible characters in the same pass and removes C2PA metadata from image files, but note that neither is the text watermark — the rewrite is what removes that.

The honest verdict

Manual editing is fine for polishing tone, but it was never designed to break a statistical watermark and rarely does. A structural rewrite through a non-watermarking model targets the one thing that carries the mark. Neither approach can promise you will pass a future AI detector — Anthropic keeps its detector private and style-based checkers are inconsistent — but the rewrite is the only one that reliably breaks the watermark itself.

References

• Anthropic: Claude text watermark announcement
• Nature: Scalable watermarking for identifying LLM outputs (SynthID-Text)
• Kirchenbauer et al., "A Watermark for Large Language Models"
• C2PA: Content Credentials Specification