Google has published SynthID for images, audio and video. Whether Gemini's text output carries a production watermark is not something we can confirm, and we are not going to assert it either way. Google has published research and tooling on watermarking AI-generated text, but we have not found a public statement that Gemini responses are watermarked in production, and a claim like that deserves a source rather than a guess.
What we can tell you is what is actually in your file. Whatever invisible characters you find in Gemini output arrived the same way they arrive everywhere else — through the browser, the clipboard, or whatever you pasted it into. Those are real bytes, you can find them, and you can delete them. Run the cleaner or list them yourself.
Every “Gemini watermark remover” page that promises to strip a Google watermark out of your text is selling the same thing the Claude pages sell: a character cleaner dressed up as a watermark remover.
SynthID is Google's approach to marking AI-generated media. The part that matters for this page is the boundary: the published deployments are on images, audio and video. If Gemini generated an image, that image may carry a SynthID mark. That is a different object from the text of a chat reply, and the two should not be treated as the same claim.
Google has also published work on watermarking AI-generated text. Published research is not the same thing as a shipped product setting, and we have not found a public statement saying Gemini replies are watermarked today. Where we cannot point at a source, we say so rather than filling the gap.
Because the characters are not a watermark. They are artefacts of the trip your text took:
None of that says anything about who wrote the words. It says something invisible is in your file, which is worth knowing and easy to fix. How to list every one →
Before you bulk-delete: zero-width non-joiner (U+200C) and zero-width joiner (U+200D) are required in Arabic, Persian, Hindi and several other writing systems. Directional marks are required in bidirectional text. A blanket strip will corrupt those files. Scan first, look at the code points you actually found, then decide.
What nothing can remove is a statistical watermark, if one is present: that is a property of how the tokens were chosen, not bytes in your file. The same limit applies to Claude, and the reasoning is identical.
We have not run anything against Gemini itself to test for a text watermark, and we are not going to imply we have. This page sticks to what Google has published and to what we measured ourselves on ordinary files. Where a claim would need a source we do not have, the answer on this page is “unconfirmed” rather than a confident either way.
We cannot confirm that it does. Google has published SynthID for images, audio and video, and it has published work on watermarking AI text, but we have not found a public statement that Gemini replies carry a production text watermark. Treat any site that answers this with a flat yes or no as guessing, unless it can point at the source.
SynthID on media is a different kind of mark from anything in this page's text discussion, and we have not tested removal on generated media, so we will not claim a method works. What we can tell you is the part we did measure: the file's metadata blocks — EXIF, XMP, provenance credentials — are removable without re-encoding the image, and that is shown with real byte counts here.
Because copying through a browser picks up characters the page was using, not because a watermark was inserted. Zero-width joiners and non-joiners, directional marks and word joiners all occur naturally in web content and in right-to-left scripts. They are bytes in your file and you can remove them — but check which code points they are first, because some of them are required by the languages those files are written in.
The character-cleaning part is the same, because the characters are not specific to any model. The watermarking part is not comparable: Anthropic has published that Claude's text output is watermarked and that nothing is added to the text, while for Gemini we have no equivalent published statement about text. So the honest versions of those two pages cannot say the same things, whatever the tool sites claim.
No. Detectors read writing patterns, not invisible characters, so deleting the characters leaves their score unchanged. That is true for Claude output — Anthropic calls the two methods “fundamentally different” — and the same separation applies here. The longer answer →