Can SynthID Be Removed? Text vs Images, Honestly
Short answer: two answers. SynthID Text lives in word choice, so a thorough rewrite or a translation through another model removes it, light edits only weaken it, and nobody outside Google’s waitlist-only portal can verify either way. SynthID Image lives in the pixel values and is designed to survive screenshots, re-encoding, resizing and cropping; what a remover strips is the C2PA metadata beside it, which is a different signal, and the pixel mark stays.
- Type
- Statistical
- Hidden characters
- No
- Official detector
- Partial: Partial
- Locally removable
- No
"Can SynthID be removed?" is two questions wearing one coat. For text the answer is yes, by rewriting, and nobody can prove it. For images the answer is no, not in any way you can verify, and the thing that does come off is not SynthID. Both halves below, with the mechanism behind each.
Text: yes, by rewriting, and only that
SynthID Text lives in word choice. A secret key tilts the model toward certain next words among the ones it considered equally good, and a detector holding the key measures the tilt across the passage. So the mark is the words. Change enough of them and the mark is gone; change a few and it is weakened but still there.
What removes it: a thorough rewrite, sentence by sentence, by a person or by a model that is not Gemini. Translation through a different model does the same, because the translated words were chosen by something without the key. Google's own documentation says the signal is less reliable on text that has been thoroughly rewritten or translated.
What does not remove it: stripping characters, because there are none to strip; swapping a few synonyms; reformatting; copying and pasting; running a "humanizer" that changes a fraction of the wording. Does editing remove an AI watermark? covers the threshold in detail, and it is high.
The catch is verification. Google's SynthID Detector portal is waitlist-only, so nobody outside it can check the original for the mark or the rewrite for its absence. A full rewrite is a well-founded expectation, not a verified result. DetectMark's rewrite through OpenAI is offered on those terms: it shows how much wording changed and makes no claim about what that removed.
Images: no, and the thing that comes off is not SynthID
SynthID Image is a pattern in the pixel values, designed to survive screenshots, JPEG compression, resizing, cropping, colour changes and filters. That is the whole point of putting a mark in the pixels rather than the metadata, and what SynthID is explains the four variants.
What does come off is C2PA Content Credentials, the signed manifest Google and OpenAI ship in the same file. It is metadata, so any metadata pass removes it: a screenshot, a social upload, or the image metadata remover, which strips it and confirms the removal by re-scanning the export. Removing the manifest does nothing to the pixel mark. A tool that strips C2PA and calls the image "SynthID-free" is describing a different signal.
Could the pixel mark be attacked? In principle, and the research says so. Regeneration attacks that pass an image back through a diffusion model can degrade many image watermarks, and a 2023 impossibility result argues no strong watermark can survive a determined adversary with enough compute. None of that helps in practice, for two reasons. Google's production scheme and key are not public, so nobody can measure whether an attack worked on a Gemini image. And OpenAI, which does run a public verifier for its own SynthID, forbids using it to test removal. You would be degrading a picture to maybe defeat a mark you cannot check for.
That is why the SynthID remover page has no button. A button implies a verdict, and there is none to give. The one image whose SynthID a browser can check today is a ChatGPT one, through the OpenAI image verifier; how to check for SynthID lists every other route and who qualifies for it.
OpenAI images: the same mark, one extra rule
Since May 19, 2026, ChatGPT images carry SynthID under licence from Google, and OpenAI's Content Provenance API can verify it. DetectMark runs that check on consent, on the original file only. It never sends a cleaned copy, because OpenAI's terms forbid using the verifier to test a removal and because DetectMark's receipts already say the pixel mark is not verified locally. The verify flow explains what a result proves.
Audio and video
The same split applies. SynthID Audio is in the waveform and SynthID Video is in the frames; a rewrite has no equivalent, and a re-encode does not remove them. DetectMark does not process either, so the honest verdict on this site is Unsupported.
The short version
| SynthID Text | SynthID Image | |
|---|---|---|
| Where the mark lives | Word choice | Pixel values |
| Removed by | A thorough rewrite or translation through another model | Nothing you can verify |
| Not removed by | Character stripping, light edits, copy-paste | Screenshots, re-encoding, resizing, cropping, metadata stripping |
| What a remover tool actually removes | Nothing | C2PA metadata, which is a different signal |
| Can the result be verified | Not outside Google's portal | Not outside Google, and OpenAI forbids testing removal |
Everything on this page also applies to Claude's watermark, which uses the same technique as SynthID Text, and to what DetectMark cannot remove in general.
See what actually comes off an image
Metadata removed and verified · pixels untouched
FAQ
Yes, by a thorough rewrite through a person or a model that is not Gemini, or by translation through another model. Character stripping, synonym swaps and light edits do not remove it. Nobody outside Google’s portal can verify the result.
Not in any way you can verify. The mark is in the pixel values and survives screenshots, JPEG re-encoding, resizing, cropping and filters. Tools that claim to remove it strip the C2PA metadata, which is a separate signal.
No. A screenshot removes the Content Credentials, which are metadata, and leaves the pixel mark, which SynthID is designed to keep through exactly that kind of transformation.
Because the outcome could not be verified. Google’s scheme and key are not public, and OpenAI forbids using its verifier to test removal. A button that reports a result nothing checked would be a lie with a progress bar.
Yes. Claude’s mark uses the same technique as SynthID Text, so the same rule holds: a full rewrite removes it, nothing less does, and no public detector can confirm it either way.
- 1. Google DeepMind: SynthID
- 2. Google: SynthID safeguards documentation
- 3. Dathathri et al.: “Scalable watermarking for identifying large language model outputs”, Nature 634 (2024)
- 4. Krishna et al.: “Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense” (arXiv, March 2023)
- 5. Zhang et al.: “Watermarks in the Sand: Impossibility of Strong Watermarking for Generative Models” (arXiv, November 2023)
- 6. OpenAI Help Center: “Provenance signals (Content Credentials, SynthID) in OpenAI-generated content” (updated August 2026; states the goal of expanding provenance signals to text)