Updated: September 3, 2026 · Docs · Clothoff AI Editorial Team
What is SynthID?
A glossary entry on the watermark Google DeepMind embeds in generated media, how it is detected, and where it stops being useful.
Educational glossary entry by a review publisher; not legal advice.
Which media types does SynthID cover?
SynthID is a watermarking system from Google DeepMind that embeds an imperceptible signal into AI-generated images, video, audio and text so the output can later be identified as synthetic. Introduced for images in 2023 and extended to video, audio and text, it is designed to survive cropping, filters and compression, and can be checked through a detector.
Google applies the watermark differently per medium. The table lists where the signal lives and how a reader can check for it.
| Medium | Where the signal is embedded | How to check |
|---|---|---|
| Images (Imagen family) | Pixel-level pattern invisible to the eye | Upload to Gemini or the SynthID Detector portal |
| Video (Veo) | Per-frame pixel pattern | Same detector; frames analyzed |
| Audio (Lyria, NotebookLM) | Spectrogram-domain pattern | Detector portal |
| Text (Gemini) | Adjusted token probabilities during generation | Open-source text detector |
How does SynthID work?
For images, a neural network modifies pixel values in a pattern the eye cannot see, and a paired network later scores whether the pattern is present. Because the signal is spread across the picture, a crop, a color filter or JPEG compression removes only part of it and detection usually still succeeds.
For text the approach is different: during generation the model slightly favors certain tokens, producing a statistical signature that a detector can score across a passage. Short texts carry less signal, so confidence falls with length.
What are its limits?
SynthID identifies output from Google models only. An image from another vendor’s generator carries no SynthID signal, so a “not detected” result does not mean a picture is real. The detector reports a likelihood, not a certificate, and heavy regeneration or re-drawing can erase the mark.
Google states the method is not a silver bullet and pairs it with metadata. SynthID therefore works best alongside C2PA Content Credentials: the watermark survives metadata stripping, while the credential carries tool and edit details.
Why does it matter for the tools we review?
None of the undress-focused services in our catalog uses Google’s image models, so SynthID does not appear in their output. Its relevance is comparative: it shows what a robust invisible watermark looks like when a vendor commits to labeling.
Video generators covered in our docs, including Veo, do carry the mark. When a reviewed tool wraps a third-party model, we check whether the vendor’s watermark survives the tool’s export path or is stripped along the way.
“SynthID watermarks and identifies AI-generated content by embedding digital watermarks directly into AI-generated images, audio, text or video.”
— Google DeepMind, SynthID overview page, accessed September 3, 2026
Consent Note
This site is a publisher: it reviews third-party tools and does not create, edit, accept or store images. Docs entries are general information, not legal advice.
Sources
All sources accessed on September 3, 2026; educational entry, not legal advice.
FAQ: SynthID
Can I check any image for SynthID?
You can upload an image to Gemini and ask whether it was generated by Google AI, or use the SynthID Detector portal where access is available. The result only speaks to Google models. An image from another generator returns no signal, which is not evidence that the picture is authentic.
Does SynthID survive screenshots and compression?
Google designed the image watermark to survive common transformations such as cropping, resizing, color filters and lossy compression, and screenshots usually preserve it because the pattern is in the pixels. Heavy edits, repainting or generating a new image from the original can remove it.
Is SynthID visible in the image?
No. The signal is imperceptible to viewers and does not change how the picture looks. That is the point: a visible logo can be cropped out, while a pixel-level pattern spread across the frame requires deliberate destruction of the image to remove.
Does SynthID work on text and audio?
Yes. For text, Gemini adjusts token probabilities during generation to create a statistical signature that an open-source detector scores. For audio from Lyria and NotebookLM, the mark is embedded in the spectrogram. Confidence for text drops with very short passages.
Do undress apps use SynthID?
Not in our testing. The services we review do not run Google image models, so no SynthID signal is present in their output. Some general platforms that expose Google video models pass the mark through; we record whether it survives the export path in the relevant review.
Is SynthID the same as Content Credentials?
No. SynthID is an invisible watermark inside the media itself. Content Credentials are signed metadata defined by the C2PA standard and attached to the file. Google uses both where possible, since the watermark survives metadata stripping and the credential carries edit history.