How to Remove Gemini’s SynthID Text Watermark (I Tested 5 Methods)

When Anthropic flipped on Claude’s text watermark in August, my inbox filled up with the same question: does Gemini do this too?

Short answer: yes, and it has for a while. Google expanded SynthID to text generated by the Gemini app and web experience back in 2024, long before the EU AI Act made it a compliance issue. Most people never noticed, because that’s the whole point of the design. Then Article 50 kicked in on August 2, everyone started auditing their AI stack, and a two-year-old feature became this month’s panic.

So if you draft with Gemini, even just outlines or rough edits, your text may carry a statistical fingerprint. Not metadata. Not hidden characters. The fingerprint lives in the word choices.

The obvious next question is how to remove it. That’s the wrong question, and I’ll get to why. But I tested five methods, because SynthID has quirks that Claude’s watermark doesn’t have, and what you learn poking at them is more useful than the removal itself.

How SynthID’s text watermark actually works

Google published the method in Nature and open-sourced the implementation, so none of this is speculation.

Language models generate text one token at a time, and each candidate word gets a probability score. SynthID nudges those scores during generation using a technique called tournament sampling, keyed to a secret value. The output reads completely normal, but across enough words the pattern becomes statistically checkable if you hold the key. Google’s developer documentation describes it as a logits processor applied after Top-K and Top-P sampling.

Three design details matter for everything below:

1. It needs choice. The watermark lives in the freedom to pick between near-equivalent words. Google’s own docs admit it is “less effective on factual responses” because there’s less room to vary word choice without hurting accuracy.
2. It skips repetition. Repeated n-grams in the recent context are deliberately not watermarked. Researchers at ETH Zurich confirmed this skip rule experimentally: feed the system repetitive text and whole stretches carry no signal at all.
3. It’s brittle by admission. Google’s docs state the watermark “can be greatly reduced when an AI-generated text is thoroughly rewritten, or translated to another language”. That’s the vendor talking, not a hacker forum.

One more thing: SynthID only detects Google’s models. It can’t see Claude’s mark, Anthropic’s detector can’t see SynthID, and there is no cross-vendor detector. Every provider is its own statistical universe.

5 ways to remove Gemini’s SynthID watermark

Quick summary, then the detail.

1. The low-entropy trick: exploit the skip rule and the factual-content gap
2. Switching to the API: the myth that won’t die
3. The translation loop: Google concedes this one themselves
4. Cross-model rewriting: works, with a new catch
5. Multi-model workspace: what I actually do (and the only one that improved the writing)

1. The low-entropy trick

This one is unique to SynthID, and it comes straight from the design details above.

Because the watermark skips repeated n-grams and weakens on factual, low-variability content, you can starve it of places to hide. Ask Gemini for bullet lists instead of prose. Ask for structured, factual output. Keep passages short, since detection confidence grows with length. The ETH team showed that repetitive input trips the skip rule so consistently that their detection queries came back carrying no key information whatsoever.

Does it work? Partially, and that’s the problem. You’re not removing a watermark, you’re contorting your writing into the shapes the watermark can’t grip. Fine for a data summary. Useless for anything where the prose is the point, which is presumably why you’re worried about a text watermark in the first place.

2. Switching to the API

The most repeated advice in every forum thread: “just use AI Studio or the Gemini API, the watermark only applies to the app”.

Here’s what’s actually verifiable. Google’s official materials scope text watermarking to “the Gemini app and web experience”, full stop. Nobody at Google has publicly said the API is exempt. A developer asked exactly this question on Google’s own forum while trying to meet their Article 50 obligations, and the thread sat unanswered. And independent testing by Karen Spinner, who built detectors against the open-source SynthID implementation, concluded the API most likely watermarks its text too.

So the popular workaround rests entirely on Google’s silence. If your plan for avoiding detection is “the vendor probably forgot to mark this surface”, you don’t have a plan. You have a hope with an expiration date.

3. The translation loop

Same trick people use on Claude’s watermark: generate in German or Japanese, then round-trip back to English through DeepL or Google Translate.

It probably works even better on SynthID, since translation resistance is a weakness Google documents openly. The pattern lives in English token choices, and a round trip rebuilds every sentence from scratch.

The cost is the usual one. Machine-translated English reads like machine-translated English. Idioms flatten, rhythm dies, and you’ll spend twenty minutes making it sound like you again. Somewhere in that editing pass you’ll realize you’re doing method 5 the hard way.

4. Cross-model rewriting

The structurally sound option. SynthID’s pattern is specific to Google’s key and Google’s token distributions. Run the text through a different model and it regenerates from a different probability universe entirely. Academic work this year has shown tournament-sampling watermarks can be broken with targeted modifications, and a full cross-model rewrite is the bluntest version of that attack.

One catch that didn’t exist a year ago: the second model may stamp its own mark. Rewrite Gemini text with Claude and you’ve swapped SynthID for Anthropic’s watermark, since Claude now marks everything it outputs. As of late August 2026, ChatGPT text watermarking is still not confirmed shipped, so GPT is currently the cleaner rewriter. That will change; OpenAI endorsed the same EU code and reportedly built the tooling years ago.

Cross-model rewriting doesn’t remove watermarks so much as let you choose whose watermark you wear. I covered Claude’s version of this problem in detail here. Which brings us to the actual point.

5. What I actually do: stop treating any single model’s output as final

Somewhere between testing methods 2 and 4, the framing flipped for me.

Every method above treats the AI draft as the finished product and the watermark as a stain to scrub off. But the reason watermark detection scares people is that they’re shipping raw model output. That was a problem before any watermark existed. Raw output has a voice, and it isn’t yours.

My workflow for anything that matters: draft with one model, then hand the draft to a different model with a different persona and ask it to challenge the argument, not the wording. Gemini drafts, Claude pushes back on logic, I rewrite the result in my own words with both critiques open next to me. I run this in HaloMate because it lets me switch models mid-conversation without rebuilding context, but the principle works anywhere you can get a genuine second opinion.

The text that comes out the other end isn’t watermark-free because I scrubbed it. It’s watermark-free because by the time I’m done, no single model wrote it. Google’s own documentation says a thorough rewrite defeats the mark, and a thorough rewrite is just another name for doing the writing. The statistics have nothing to grab, and, not coincidentally, it’s better prose than any of the other four methods produced.

FAQ

Does Gemini watermark its text?

Yes. Google expanded SynthID to text generated through the Gemini app and web experience in 2024, and it applies today. The watermark is embedded by adjusting token probabilities during generation and is invisible to readers.

Does the Gemini API watermark text too?

Officially unconfirmed. Google’s materials only name the app and web experience, and Google has not answered direct questions about the API. Independent testing suggests the API most likely watermarks text as well. Treat API output as marked until Google says otherwise.

Can I detect SynthID myself with the open-source code?

Not on Google’s text. The open-source SynthID-Text release lets you watermark your own model’s output with your own key. Detecting Google’s watermark requires Google’s key, which is not public. The open code tells you how the lock works, not where the key is.

Does paraphrasing remove the SynthID text watermark?

Partially. Google’s own docs say the signal is greatly reduced by thorough rewriting or translation, and weaker on factual or repetitive text to begin with. Light synonym-swapping is unreliable; a full rewrite removes it entirely, at which point the text is arguably no longer AI-generated.

Is removing an AI watermark illegal?

The EU AI Act’s marking obligations fall on providers like Google and Anthropic, not on end users. Removing a watermark from your own drafts isn’t a crime in itself, but presenting AI text as human work can still violate your employer’s policy, a school’s honor code, or a publication’s disclosure rules. The watermark isn’t the risk. The misrepresentation is.

The part nobody wants to hear

Gemini’s watermark has been there for two years. You didn’t notice, your readers didn’t notice, and nothing bad happened, because detection requires someone to actively check with the right vendor’s key. The panic isn’t really about the watermark. It’s about the sudden realization of how much unedited model output we’ve all been shipping.

The providers will keep rotating: Claude marks today, Gemini has marked since 2024, ChatGPT will mark eventually. Chasing removal methods means re-fighting this battle every time a vendor rotates a key.

Making the words your own means never fighting it again.