The watermark is statistical
SynthID-Text influences token selection probabilities. There are no hidden characters to delete and no metadata to strip.
Parity-Preserving Anthropic SynthID-Text Watermark Scrubber
A clean-room paraphrasing pass that changes statistical authorship without flattening your ideas, facts, or formatting.
Open the scrubberPaste Claude-generated text
Meaning intact. Pattern replaced.
Your rewrite will appear here
Code fences return untouched. Structure and facts stay put.
No text is stored by Unthropic. Requests are processed in real time.
DEEPSEEK V4 FLASH / LUNA FALLBACK
Why paraphrasing works
SynthID-Text influences token selection probabilities. There are no hidden characters to delete and no metadata to strip.
A full rewrite through a non-Claude model samples a different token path while retaining the original claims and structure.
Fenced code is masked before the request and restored verbatim afterward, keeping syntax and variables out of the rewrite path.
Common questions
Because the signal lives in the probability distribution across token choices, not in a detectable string or character sequence.
Rewriting with a model from a different family avoids reproducing Claude’s token-selection signature. The goal is a genuinely independent generation path.
The rewrite prompt preserves Markdown structure, while fenced code blocks are removed from model context and restored byte-for-byte after rewriting.