NewOctober 5, 2026: OpenAI announces text watermarking for ChatGPT models. Learn more →
Clean AI WatermarkSynthID-Text is Google DeepMind's word-pattern watermark, used by Gemini since 2024 and adopted by Claude in 2026. Paste the text below to break the pattern.
Paste on the left and hit Remove watermark. Takes about ten seconds.Paste above and hit Remove watermark. Takes about ten seconds.
As the model writes, SynthID-Text scores the candidate words for each position with a secret key and nudges the model toward the higher-scoring ones. Each choice is tiny, but across a few hundred words the tilt becomes a statistical fingerprint. A detector with the key sums the scores and returns a verdict. Google DeepMind published the method in Nature in October 2024 and open-sourced the code; Anthropic's Claude watermark is a version of it.
The fingerprint depends on the steered words staying in place. We swap enough of them for equivalents to collapse the signal, and few enough that your sentences, tone and length stay as they were. Hidden characters and markdown leftovers go in the same pass.
Gemini since 2024, and Claude models launched since August 2026, with older Claude models being added. The scheme is open source, so other providers can adopt it.
A translation by a person does. A translation by a watermarked model adds a fresh watermark, because every word in it is the model's.
Detection needs the provider's key. Google and Anthropic hold theirs and grant access to eligible organizations, so a random reader cannot run the check.