How to remove Claude's watermark
Five methods compared with a plain verdict on each. Stripping invisible characters: no effect, because Claude inserts none. Light editing: minimal effect, per Anthropic's own answer. Local synonym rewriting: partial effect, measurable only as the percentage of wording changed. Open-source cleaners such as claude-watermark-cleaner: genuine effect, running rewrite passes until roughly seventy per cent of token sequences are disrupted. Full rewriting through a non-Claude model: most effective, and also the most work.
The honest summary is not that nothing works. Rewriting genuinely works, because the mark is carried by word sequences. What nobody can do is verify the result, because the detector is private.
Frequently asked questions
What actually removes Claude's watermark?
A substantial rewrite. Because the mark is carried by the sequence of word choices, changing most of those choices disrupts it. Character-level cleanup does nothing at all.
Do watermark remover tools work?
Tools that only strip invisible characters do nothing to a statistical watermark. Tools that genuinely rewrite the prose do disrupt it, though none can confirm the result.
Is removing an AI watermark legal?
This is not legal advice. Removing a mark does not remove a disclosure obligation you may have to a school, employer, client or regulator, and under the EU AI Act transparency rules that obligation exists independently of the watermark.