How to remove Gemini's watermark from text

Gemini text carries SynthID, the longest-running text watermark in production. What is in Gemini output, what can be stripped exactly, and what only a rewrite touches.

Removing the watermark · 3 min read

Gemini has carried a text watermark longer than any other major model. Google shipped SynthID-Text into production and published the mechanism in Nature in October 2024, well before the EU AI Act made marking a compliance requirement for everyone else.

That head start means Gemini output is the best-understood case, and also the one where the honest answer is most constrained.

Two things are in Gemini output

Characters you can see and count. Em dashes used as mid-sentence connectors, curly quotes, the single-codepoint ellipsis, non-breaking spaces, occasionally zero-width characters. These are real codepoints. They can be matched exactly and deleted, and you can confirm the result by re-scanning.

SynthID, which you cannot count. A statistical bias in which tokens the model selected. Nothing is inserted into the text. There is no codepoint to find.

Most guides to removing a Gemini watermark only address the first, then claim to have addressed the second. They are different problems with different answers.

The character layer

Straightforward and provable:

Scan, strip, re-scan, confirm zero. No model involved and nothing about your wording changes.

Worth doing regardless of the watermark question. These characters break search, diffs, code blocks and CSV imports, and they are what a human editor notices first.

The SynthID layer

Here is what Google's own paper documents as degrading the signal:

• paraphrasing
• heavy editing
• translation
• short passages, which never carry enough signal to detect at all

The reason is structural. SynthID biases token selection among near-equivalent candidates, so the evidence is the specific words that were chosen. Change the words and the evidence goes. Google published this limitation themselves; it is not an outside claim.

But nobody outside Google can check the result. Detection needs Google's production key. The algorithm is open source; the key is not. So no third-party tool can tell you whether Gemini's mark is present before you start or absent after you finish.

Any service that reports a Gemini watermark percentage is running a generic AI-detector heuristic and calling it a watermark reading. Those are two different things, and the heuristic produces false positives on non-native English writers, technical writing and heavily edited prose.

What actually helps

Length works against you. SynthID needs a few hundred tokens to accumulate detectable signal. A long document carries a stronger pattern than a short one. If you are working with a short passage, it was never reliably detectable to begin with.

Mixing dilutes. Text where Gemini output is interleaved with your own writing carries a weaker signal, because your sentences contribute nothing to the pattern.

Rewriting is the only lever on this layer. Not reformatting, not character stripping, not changing the font. Word choices are the carrier, so word choices are what has to change.

What we do, precisely

The character strip is exact and verifiable — re-scan the output and every count is zero.

The rewrite changes word choices throughout while holding facts, structure, register and length steady. Paraphrase is the mechanism Google's paper names. Whether a given rewrite degrades the signal enough is unmeasurable without Google's key, so we do not put a number on it.

That is less satisfying than "100% removed" and it is the only version that is true.

Run a free scan to see the character layer in your Gemini text right now.

Related: can you remove SynthID? · what SynthID-Text is · Claude vs ChatGPT vs Gemini · removing ChatGPT's watermark