How to Remove SynthID Watermarks from AI Text

Google's SynthID can mark AI-generated text in a way that's completely invisible to the human eye. You can't see it, you can't feel it, and most third-party AI detectors won't find it. But Google's own system can, because the watermark is baked into how the text was generated in the first place. If you're using Gemini to draft content and need to clean that signal out, you need to understand what you're working with before you can remove it.

The practical news: SynthID watermarks are surprisingly fragile. Substantial rewriting breaks them.

Removing a SynthID watermark means rewriting the text thoroughly enough that the original token patterns break down. SynthID is embedded during generation, not added afterward. An AI humanizer that restructures at the sentence level, changing word choices and syntax throughout, disrupts the signal in a single pass. Around 30-50% of tokens need to change for detection to fail.

What Is SynthID and How Does It Work?

SynthID isn't a tag you can strip, a hidden character you can delete, or metadata you can remove in a text editor. It works earlier than that, inside the model itself.

Google DeepMind's SynthID text watermark, introduced in 2024, works by manipulating the token selection process during text generation rather than adding metadata after the fact. When a large language model generates text, it samples from a probability distribution at each step. SynthID biases this sampling with a secret key that shifts which tokens get chosen. The resulting text looks identical to a human reader, but the watermark detector can identify the pattern across thousands of tokens. Google published the technical approach in a 2024 Nature paper co-authored by DeepMind researchers. The watermark is currently deployed in Gemini family models. A critical limitation acknowledged in that paper: the watermark degrades significantly when text is paraphrased or substantially edited. Rewriting roughly 30-50% of tokens breaks the signal enough to defeat detection in their own tests. That fragility is the mechanism you can work with.

This is fundamentally different from image watermarks. A photo watermark is overlaid on existing content. A SynthID text watermark is woven into the content as it's created, which means it only holds as long as most of the original words stay intact.

Can SynthID Watermarks Be Removed?

Yes, and the DeepMind paper itself explains why.

The watermark degrades proportionally to how much the text changes. Light edits (fixing typos, adjusting punctuation) leave the signal intact. Moderate edits (rephrasing individual sentences) weaken it. Heavy rewriting (restructuring paragraphs, changing word choices throughout) breaks it.

Google's own researchers tested several attacks on the watermark, including paraphrasing with a separate AI model. Their conclusion: watermark robustness drops substantially when around half of the tokens change. The signal doesn't survive aggressive humanization.

Two scenarios where removal is harder:

• Very short text (under 200 words) doesn't have enough tokens to form a clear watermark pattern, which also means partial edits don't disrupt it as cleanly
• Light paraphrasing (swapping synonyms, shuffling a few clauses) leaves the core structure intact, and so does most of the watermark

The practical takeaway: you need a rewrite that genuinely transforms sentence structure and word choices, not just a surface polish. If you want to understand the broader picture of how to make AI text undetectable, the same principle applies: depth of change is what matters, not cosmetic edits.

How Rewriting Removes the SynthID Signal

SynthID isn't checking whether your text "sounds" AI-generated. It's checking for a specific statistical fingerprint in the sequence of tokens. When you rewrite, you're producing a new sequence, one the SynthID system has no record of.

Think of it like a song recorded in a specific key. Transpose it to a different key and it still sounds similar, but the acoustic fingerprint has changed. Rewrite AI text substantially and the token fingerprint changes with it.

A few approaches that work:

Full sentence restructuring. Changing clause order, splitting long sentences into two, merging short ones: each of these changes the token sequence at that position. Do this across every paragraph and the cumulative effect breaks the watermark.

Tone shifting. Running text through an Academic or Professional tone mode in an AI humanizer doesn't just rephrase words. It changes the syntactic patterns the original model used, so token choices diverge from the original early in each sentence.

Building from notes. If you used Gemini to draft something, copy the key points into a new document and rewrite from scratch using them as reference. The original generation never appears in your final text at all.

Combining methods. Running through a humanizer first, then doing a manual pass on paragraphs that still feel close to the original, gives you the deepest disruption.

The process is almost identical to how to humanize AI text for any AI detector bypass. SynthID is just a watermark system rather than a pattern-recognition detector, but the fix is the same: change the text enough that the original token sequence doesn't survive.

How to Use NaturalRewrite to Remove the SynthID Watermark

NaturalRewrite's humanization pipeline does sentence-level restructuring, not just synonym swapping. That depth of rewriting is what makes the SynthID signal degrade.

Here's the workflow:

1. Paste your Gemini-generated text into NaturalRewrite. The tool accepts up to 300 words on the free tier, 1,500 on Starter, and 3,000 on Pro.

2. Choose a tone mode that differs from the original. If the Gemini draft sounds formal, try Casual. If it's casual, try Academic. The bigger the tonal shift, the more sentence-level restructuring happens. Starter and above unlock all 5 tone modes.

3. Run the humanization. NaturalRewrite's multi-model pipeline rewrites the text at the sentence and paragraph level, changing both word choices and syntactic structure.

4. Check with the built-in AI detection checker. This tests whether the humanized output reads as AI-generated to common detection models. A clean pass there also correlates with enough token disruption to break most statistical watermarks. Free tier gets 3 checks per day; Starter and above are unlimited.

5. Do a quick manual review. Read through the output. If any paragraph still feels very close to the original phrasing, rewrite that section manually before finalizing.

For longer documents, process them in chunks. Run each section through separately, then stitch the outputs together. NaturalRewrite handles one text at a time, so breaking the document into logical segments is the right approach.

What Won't Work

A few common attempts that don't remove the watermark:

Running text through a spell checker or grammar tool. These make surface-level changes only. The underlying token distribution stays nearly identical.

Asking ChatGPT to summarize and rewrite. Summaries change the length but often preserve sentence structure and vocabulary patterns. The resulting text may still carry enough of the original fingerprint to register.

Changing fonts or formatting. This affects document metadata, not text tokens. Completely unrelated to how SynthID works.

Using Gemini to rewrite its own output. If you ask Gemini to rewrite what it just wrote, you're still using the same model with the same token biases. The new output gets a fresh SynthID watermark, just a different one.

The common thread: none of these change enough tokens to break the statistical pattern. Only genuine rewriting at the sentence level does.

Frequently Asked Questions

Does SynthID work on text from ChatGPT or Claude? SynthID text watermarking is a Google DeepMind technology, currently deployed in Gemini models only. Text generated by ChatGPT, Claude, or other non-Google models doesn't carry a SynthID watermark. Those providers may have their own watermarking systems, but SynthID specifically isn't one of them.

How much of the text needs to change to remove SynthID? The DeepMind paper found the watermark degrades substantially when around 30-50% of tokens change. If you've restructured most sentences and replaced a significant portion of the vocabulary, the signal is gone. A light polish won't reach that threshold.

Will NaturalRewrite specifically detect SynthID before humanizing? NaturalRewrite's built-in detector checks whether text reads as AI-generated to common detection models. It doesn't identify SynthID patterns specifically. But the humanization process changes the text deeply enough that the original SynthID fingerprint doesn't survive.

Is removing a SynthID watermark against Google's terms of service? Google's terms for Gemini don't prohibit editing or rewriting generated content. SynthID is positioned as a detection tool for Google's internal use, not a content restriction mechanism. Standard editing and rewriting is permitted.

Does SynthID survive translation? Translation through a different model generates an entirely new token sequence in the target language. SynthID watermarks don't transfer across languages or models, so translation to another language and back would break the signal, though that approach introduces its own quality problems.

If you're working with Gemini-generated text and need to remove the SynthID signal, NaturalRewrite's humanization process gives you a clean, restructured output with the original token fingerprint stripped away. Start on the free tier to test the workflow, then move up to Starter or Pro for longer documents. Try it at naturalrewrite.com.