How to remove a SynthID watermark
SynthID-Text is Google DeepMind’s watermark for AI-generated text, applied to text from the Gemini app and web. The idea went mainstream in 2026, when Anthropic and OpenAI shipped their own statistical watermarks for Claude and ChatGPT. Here’s how the SynthID watermark works, why it’s hard to remove by editing, and how to strip it properly. You can do it free with the Unmark homepage scanner.
Updated .
How it works
When a language model writes, it repeatedly samples the next word from a probability distribution. SynthID nudges that sampling with a secret pseudo-random key so the chosen words follow a subtle statistical pattern across the whole passage. Nothing is visibly different — the text reads normally — but a detector holding the key can measure the pattern and estimate whether the text was machine-generated. Because the mark lives in the sequence of word choices, it doesn’t rely on hidden characters or metadata.
Google’s own documentation notes that detector confidence can drop sharply when text is thoroughly rewritten or translated, and that the watermark is weaker on factual answers, where the model has less freedom to vary its wording. Google’s SynthID Detector checks images, video and audio — not text.
Why editing a few words doesn’t remove it
The watermark is distributed, not localized. Swapping a word here and there, fixing punctuation, or trimming a sentence leaves most of the pattern intact — often enough for a detector to still flag it. The only reliable way to remove a statistical watermark is to change enough of the token sequence that the pattern no longer holds, i.e. a genuine rewrite.
We measured it. In our test of 120 passages watermarked with the open-source SynthID-Text, swapping 10% of the words for synonyms left 92% still detectable; after Unmark’s default rewrite, 3% were.
How to remove it
- Paste the text into the Unmark homepage scanner.
- Choose a rewrite strength and remove the watermark. Unmark rewrites the wording to disrupt the statistical pattern while preserving your meaning, and strips any invisible characters in the same pass.
- Copy the clean text and re-read it to confirm the meaning is intact.
A realistic expectation
Rewriting substantially degrades SynthID and other detector signals, and heavier rewrites degrade them more — but no tool can guarantee text is permanently undetectable against every detector. Treat watermark removal as best-effort, and always review the result.
60 free rewrites an hour, no sign-up.