SynthID Explained: How Google Watermarks Text, Images, Audio and Video
Updated September 2026
SynthID is one name for several different watermarks, and the differences decide what you can do about them. The image, audio and video versions live in the content. The text version lives in word choice. None of them lives in metadata, which is why a metadata tool, StripShot included, cannot remove them and should not pretend to.
Quick answer
The short version:
- 1SynthID for images, audio and video is embedded in pixels or waveform. It survives cropping, compression and screenshots. No metadata tool removes it.
- 2SynthID-Text (Nature, October 2024) nudges token choices while a model writes. It survives copy and paste and is weakened only by rewriting.
- 3Both need a Google-held key to detect. Public tools cannot read them.
- 4Gemini and Imagen outputs also carry C2PA and IPTC metadata. That layer is removable, and StripShot removes it.
One brand, four watermarks
| Version | Where the mark lives | How it is read | Survives |
|---|---|---|---|
| SynthID for images (Imagen, Gemini image output) | A pattern spread across the pixels, invisible to the eye | Google's detector scores the pixels | Cropping, resizing, JPEG compression, filters, screenshots |
| SynthID for video (Veo) | In the frames | Google's detector | Re-encoding, trimming, platform transcodes |
| SynthID for audio | In the waveform | Google's detector | Compression, re-encoding |
| SynthID-Text (Gemini text) | In which tokens the model chose | A keyed scoring function over the text; the model is not needed | Copy and paste, light editing; weakened by paraphrase and short length |
How the image watermark works, roughly
Google has published the outline but not the weights. One network embeds a signal into an image during or after generation, adjusting pixel values in a way that is invisible to people and spread across the whole frame. A second network reads it back and reports a confidence level. Because the signal is redundant across the image, cropping away part of it or compressing it leaves enough behind to detect. That robustness is the point, and it is also why no legitimate metadata tool can claim to remove it: the mark is not a block of bytes you can cut out, it is the picture.
What StripShot does with a Gemini image
Gemini and Imagen images carry two things: SynthID in the pixels and C2PA plus IPTC provenance tags in the metadata. StripShot removes the metadata layer at the binary level with no re-encoding, because that is the layer platforms read to apply labels. It leaves the pixels byte for byte identical, so SynthID is untouched. We would rather say that plainly than sell a false promise. The Gemini page spells this out.
How the text watermark works
SynthID-Text, described in Nature in October 2024, is a generative watermark: it changes how the model samples each word. Using a secret key and the preceding words, it scores candidate tokens and holds a small tournament that favors higher-scoring ones, so the finished text leans toward a keyed pattern while reading as normal prose. Detection recomputes the scores from the key and the text alone and asks whether the pattern is stronger than chance. Google open-sourced a reference implementation in 2024; the production key is private, so the reference code can verify a mark you created under your own key, not Google's.
The paper is candid about limits: short outputs carry too little signal, and paraphrasing, translation and heavy editing reduce detection. It also notes that current production text watermarks work token by token because streaming demands it, which is exactly the property that keeps paraphrase effective against them.
What changed on the Gemini API
Whether a given Gemini surface marks text has shifted over time, and developer forum posts in 2026 indicated the Generative Language API no longer watermarks text output on current models. Consumer Gemini surfaces may behave differently, and Google has never made the detection key public. The practical rule: assume any Gemini text might carry a mark, and assume no outside party can tell you whether it does.
SynthID and C2PA together
Google attaches C2PA Content Credentials to many outputs alongside SynthID, and the C2PA specification now describes soft bindings: an imperceptible watermark that can point back to a copy of the manifest stored elsewhere, so provenance survives even after the file metadata is gone. For images, SynthID is that soft binding. For text, the statistical mark plays the same role. Metadata stripping addresses the manifest inside the file, which is what triggers platform labels today; it does not address the soft layer, and the honest description of any cleaning tool says which layer it works on.
What you can do about each layer
- Metadata (C2PA, IPTC, XMP, EXIF). Removable, verifiable, and StripShot does it without re-encoding. This is what stops most metadata-triggered labels.
- Pixel, audio and video watermarks. Out of scope for metadata tools. Research regeneration methods exist, they degrade the content, and nothing outside Google confirms the result.
- Text watermarks. Only rewriting changes them, with no confirmation possible. StripShot does not rewrite text. It cleans hidden characters and shows writing tells, both of which are verifiable.
- Disclosure. Where you are required to disclose AI use, the EU AI Act among other rules, a removed manifest does not remove the requirement.
Check a Gemini or Imagen image
See the C2PA and IPTC tags a Gemini image carries
Runs in your browser. Nothing is uploaded.
Need unlimited strips?
Scans are free. Pro is unlimited.
Images, video, audio, documents, PDF and text in one tool. Pro adds unlimited strips, whole-folder batches and zip download with a JSON report.
Frequently asked questions
What is SynthID?
Google DeepMind's family of watermarks for AI-generated content. The image, audio and video versions embed an imperceptible signal in the pixels or waveform. The text version, SynthID-Text, adjusts which words the model picks while writing. All of them are read back with a Google-held key.
Can a metadata remover strip SynthID from an image?
No. SynthID for images is in the pixels, not in EXIF or C2PA. It is designed to survive cropping, compression and screenshots. StripShot removes the C2PA and IPTC metadata that Gemini and Imagen images also carry, and says clearly that the pixel watermark stays.
Can a text tool remove SynthID-Text?
No. SynthID-Text is in the choice of words. Character cleaning does not change words. Paraphrasing weakens the mark, and Google's paper says so, but no outside tool can confirm the result without Google's key.
Does Gemini still watermark text?
Google's public statements about which surfaces mark text have changed over time, and the detection key has never been public. Treat any Gemini text as possibly marked, and treat any claim to detect or remove the mark from outside Google as a guess.
Is SynthID the same as C2PA?
No. C2PA is a signed metadata manifest inside the file; it can be read by anyone and stripped by any metadata tool. SynthID is in the content itself and needs Google's detector. Google attaches both to many outputs, and the two are increasingly designed to work together.