Skip to content
Simple Unmark

Mechanism-specific

SynthID Watermark Remover for Text

SynthID Text encodes a signal across many token choices, so there is no single symbol to delete. Simple Unmark rewrites the passage so the wording is no longer the distribution the watermark was applied to.

Original

Local scan runs as you type.

Hidden-character cleanup + wording rewrite

How this works

Deterministic marks can be removed exactly. Probabilistic patterns can only be reduced. Simple Unmark keeps the two separate and does not promise detector outcomes.

  • The watermark is applied during generation, by biasing which of several acceptable next tokens gets chosen. Detection then tests whether a passage's choices line up with the watermarking key more often than chance would explain.
  • This has two consequences. Longer text carries more evidence, so short passages are inherently harder to detect. And highly constrained text — dense facts, code, quotations — carries less, because the model had fewer free choices to encode into.
  • Google's own documentation names thorough rewriting and translation as things that greatly reduce detector confidence, and Anthropic describes a complete rewrite as what removes its SynthID-Text-style mark. Rewriting is not a trick around the mechanism; it is the mechanism's documented limit.

Three steps

  1. 1

    Paste the passage. Longer, more discursive text carries more watermark evidence than a couple of factual sentences.

  2. 2

    Run the clean. Deterministic cleanup runs first, then the rewrite changes syntax, sentence boundaries, clause order, transitions and vocabulary.

  3. 3

    Verify the content survived: numbers, names, quotations and claims. Then copy the result.

Worked example

Before

A watermarked passage: the model's own clause order, transitions and word choices.

After

An independent passage: same claims, different structure and vocabulary throughout.

A synonym-swap is not what this does and would not work — swapping a few words leaves most of the original choices, and therefore most of the evidence, intact. Output varies per request, so no fixed sample is shown.

Scope

What this page can and cannot do

Can

  • Replace the token-choice distribution the watermark was applied to, across the whole passage
  • Explain, with citations, why character cleanup cannot address this class of watermark
  • Tell you honestly which factors make a signal weaker or stronger

Cannot

  • Score, detect or verify a SynthID watermark — no public detector for arbitrary text is available to us
  • Guarantee an outcome against an unpublished or future detector
  • Promise anything for very short inputs, where both watermarking and rewriting have less to work with

Submitted text and cleaned output are not stored in our application database, and no text is ever sent to analytics. Deterministic inspection on the invisible character remover runs entirely in your browser. See the privacy policy and methodology.

Questions

About synthid watermark remover for text

Sources

Can Unicode cleanup remove SynthID Text?
No, and any tool claiming otherwise has confused two mechanisms. SynthID Text adds no characters. Character cleanup and statistical rewriting address genuinely different things, which is why every clean here runs both.
Is neutralisation guaranteed?
No. Watermark detection is statistical and the methods evolve. The honest claim is a large reduction in the signal, and we will not claim more than the providers' own documentation supports.
Why does cleaning cost credits?
Every request includes a model-backed, meaning-preserving rewrite, which has a real per-request cost. Deterministic character inspection is free and unlimited on the invisible character remover.
Does text length matter?
Yes, in both directions. Statistical watermarks need enough text to be detected at all, and a rewrite has more room to change structure in longer passages.

Other cleaners

Need more than three cleans?

Credits cost 0.1 per started 100 words, never expire, and there is no subscription. An account raises the limit to 5,000 words per clean.

See pricing