Does Gemini watermark its text?
Yes. Google applies SynthID to text generated by the Gemini app and web experience.
SynthID adjusts the probability scores the model assigns to candidate next words, so the choices encode a signature that a detector holding the key can read. It is the same published method Anthropic uses for Claude, from the DeepMind paper in Nature. Google's own suggestion for checking a passage is to paste it into Gemini and ask whether it carries a SynthID watermark.
What Clipboard Sanitizer can do about it
The same as for Claude: nothing about the watermark, everything about the characters. Gemini output arrives with the usual typographic punctuation and, like any assistant's output, can carry invisible characters picked up along the way.
Questions
Can I check whether a passage carries a SynthID watermark?
For text, Google's route is to paste it into Gemini and ask. The standalone SynthID Detector portal is documented for image, video and audio.
Does paraphrasing remove it?
It weakens it. The paper's authors say generative watermarks are weakened by edits such as paraphrasing, and are explicit that watermarking is not a complete solution to detecting AI text.