What is SynthID? Google's invisible AI watermark, explained
SynthID is the invisible watermark Google adds to content made with its AI. How it works, how to check an image for it, and why metadata tools cannot see it.
On this page
What is SynthID?
SynthID
SynthID is a watermarking technology from Google DeepMind. It embeds an imperceptible signal in content generated by Google’s AI models (images, video, audio and text) at the moment the content is created, and a matching detector can later recognise that signal.
Google applies it across its generative products. DeepMind lists the Gemini app, Imagen, Veo and Lyria among them, and the watermark covers audio from NotebookLM’s podcast feature too. When Google launched image verification in the Gemini app in November 2025, it said that more than 20 billion AI-generated pieces of content had been watermarked with SynthID since its introduction in 2023.
How does SynthID work?
SynthID has two halves. An embedder adds the watermark while the model generates the content, and a detector checks content later for the same pattern.
For images, the watermark is spread through the pixel values themselves. Google describes it as invisible to people and designed to stand up to cropping, filters, frame-rate changes and lossy compression. For audio, Google says it survives added noise, MP3 compression and speed changes. For text, it works on the model’s word choices, and Google has open-sourced that text version so other developers can use it.
The watermark is not a tag, a field or a block of metadata, so opening the file in a metadata viewer shows nothing about it.
How can you check an image for SynthID?
Google offers two ways, both limited to content made with Google AI.
- In the Gemini app. Upload the image and ask whether it was created or edited with Google AI. According to Gemini Apps Help, you need to be signed in with a personal account or a qualified Workspace account, one file at a time, up to 100 MB. There is a usage quota of about ten image checks in any 24 hours, with separate quotas for video and audio.
- With the SynthID Detector portal. Announced in May 2025, the portal scans images, audio, video and text for the watermark and highlights where it is most likely present. Access started with journalists, media professionals and researchers through a waitlist.
What does a SynthID result mean?
Google’s help page describes the two outcomes:
- Watermark detected: all or part of the image or video was created or edited by Google’s AI models.
- Watermark not detected: the content was not created or edited by Google AI, but it could have been made by other AI systems.
SynthID tells you whether Google AI made or edited the content. It says nothing about images from other generators, which carry their own marks or none at all.
Is SynthID the same as C2PA or metadata?
No. They are different layers of the same file.
| SynthID | C2PA Content Credentials | IPTC / XMP fields | |
|---|---|---|---|
| Where it lives | In the pixels, sound or text | In a signed metadata block | In metadata fields |
| Who can read it | Google’s detector only | Any C2PA verifier | Any metadata viewer |
| Survives metadata removal | Yes, by design | No | No |
| Used by other companies | OpenAI, for its images | Any tool that writes it | Any tool that writes it |
Google has said it plans to extend verification in Gemini to C2PA Content Credentials, so that content from outside Google’s ecosystem can be checked for its declared source. We explain the C2PA side in what is C2PA.
Does a Nano Banana image carry SynthID?
Yes, and the file says so. On 27 September 2026 we downloaded an image generated in Gemini, the app behind the Nano Banana image model, and read its metadata. Its C2PA manifest, signed by Google’s C2PA library, lists two actions: “Created by Google Generative AI.” and “Applied imperceptible SynthID watermark.”
That gives you two layers to look at. The manifest is metadata: a checker like ours reads it, and it vanishes with a screenshot or a re-upload. The SynthID watermark is in the pixels: only Google’s tools see it. When we removed the metadata, the pixels stayed identical, so the watermark was still there. The full comparison with a ChatGPT image is in AI image metadata.
Does OpenAI use SynthID too?
Yes. OpenAI’s help center says images generated with ChatGPT, Codex and the OpenAI API carry both C2PA metadata and a SynthID watermark, and that its own checker at openai.com/verify looks for them. The ChatGPT image we tested declared a c2pa.watermarked.unbound action in its manifest, which matches. Google’s Gemini check and OpenAI’s checker each look for their own company’s signals.
What should you use to check an image?
Use both layers. Run the file through our AI image checker to read what its metadata declares, including Content Credentials, and ask the Gemini app about SynthID if you suspect Google AI. Neither replaces the other, and neither is proof on its own that a picture is authentic.