Does Gemini Watermark Its Text? SynthID Explained

Yes. Since May 2024, Google has marked text from the Gemini app and Gemini on the web with SynthID, an invisible watermark built into the words the model picks. Google DeepMind announced it together with watermarking for Veo video, and later published the method in Nature and released an open-source version. Google is also unusually frank about the limits: in its own words, SynthID's confidence "can be greatly reduced when an AI-generated text is thoroughly rewritten or translated to another language".

This article covers what SynthID for text is, where it applies, what weakens it, and why you can't check Gemini text for it yourself.

What SynthID for text is

SynthID started as a watermark for AI images and grew into a family of techniques for images, audio, video and text. The text version works differently from the others, because there is no image to hide anything in.

Instead, it works on the words. As Google DeepMind describes it, a language model assigns every possible next word a probability score. SynthID adjusts those scores slightly, using a secret key, so that over a passage the chosen words form a pattern. The text reads the same, and the pattern is invisible, but a detector that holds the key can score how closely a passage matches it.

Google tested this live before writing about it. In the Nature paper published in October 2024, Google DeepMind reported a trial on nearly 20 million Gemini responses and found no meaningful difference in how users rated watermarked and unwatermarked answers.

Where Gemini's watermark applies

Google's announcement names the Gemini app and web experience. It does not mention the API, and answers from Google staff on the developer forum have been inconsistent. Until Google says it plainly, treat API text as unconfirmed either way.

Google has also signed up, along with Anthropic, Meta, Microsoft and OpenAI, to the EU's code of practice on marking AI-generated content under the EU AI Act. Both Claude and ChatGPT introduced text watermarks in 2026; Gemini's is the oldest of the three.

What weakens SynthID, according to Google

Google's May 2024 post lists the limits itself:

• Thorough rewriting. Confidence scores "can be greatly reduced when an AI-generated text is thoroughly rewritten".
• Translation. The same sentence continues: "or translated to another language".
• Factual answers. SynthID "is less effective on responses to factual prompts". When there is only one right answer, there is little room to choose between words, so there is little room to nudge.
• Length. It "works best when a language model generates longer responses". A short reply carries too few nudges to measure.

The common thread is room for choice. The watermark lives in the choices the model made between good alternatives. Remove those choices, by replacing the words or by asking a question with only one possible answer, and the pattern goes with them.

This also explains what does not weaken it. SynthID is not made of hidden characters, so tools that strip zero-width spaces or swap punctuation leave it untouched. Light edits leave most of the original word choices in place. A full rewrite in new words, which is what a humanizer does, replaces most of them.

Can you check Gemini text for SynthID?

Not today, and not for lack of tools:

• The SynthID Detector portal checks images, video and audio files. Google's own page doesn't list text as supported.
• The open-source SynthID Text library is on GitHub and in Hugging Face Transformers. It lets developers watermark and detect text from their own models, with their own keys. It cannot read Gemini's watermark, because Gemini's keys are not public.
• Third-party "SynthID checkers" for pasted text do not have Google's key either. At best they run a general AI detector, which estimates from writing style and gets things wrong.

So if someone says they found a Gemini watermark in your text, ask how. Only Google can run that check on Gemini output.

What a watermark proves, and what it doesn't

Google's own framing is that SynthID "isn't a silver bullet" for identifying AI content. A detected mark is good evidence that Gemini produced a passage. A missing one proves little: the text may be short, factual, edited, translated, or written by a model that does not use SynthID.

What this means if you write with Gemini

• Check your rules first. Schools, publishers and employers set their own rules about AI help and disclosure. A watermark does not change them, and rewriting does not either.
• Assume Gemini app text is marked. Longer, open-ended replies are where the mark is strongest.
• Small edits leave most of it. A thorough rewrite is what Google itself says reduces it, and it also tends to make the text sound like you.
• Character cleaners do nothing. SynthID is in the words.

If you want the rewrite done for you, the Gemini Watermark Remover runs our humanizer on Gemini text: it rewrites every sentence in new words, then checks the result against your draft and flags any number, name or citation that changed. You can try it on the page without an account.

The short version

Google has watermarked Gemini app and web text with SynthID since May 2024; whether API text is marked is not clearly stated. The mark is a pattern in the word choices, invisible to readers and impossible to check for yourself, because Gemini's keys are not public. Google says thorough rewriting or translation greatly reduces it, and that it is weaker on short and factual answers. Stripping characters does nothing.