SynthID: Google’s AI watermark explained
Learn what SynthID is, how Google’s invisible AI watermark works, how to check an image for it, and what it means for AI product images.
- Published
- Last updated
SynthID is Google DeepMind’s watermarking technology for AI-generated content. It embeds a signal people can’t see or hear, so detection tools can identify supported AI output. For a store making product images, the watermark identifies AI involvement. It does not confirm that the image accurately shows the product for sale.
What SynthID is
SynthID adds a detectable watermark within AI-generated images, video, audio and text. Google DeepMind’s SynthID overview explains that the signal is embedded in the content itself. An image can look ordinary while still carrying it.
For images, the watermark is added during generation. It is designed to survive common changes such as cropping and compression, although extensive changes can make detection fail. In text, SynthID adjusts the probabilities used to choose the next token, a small piece of text. The resulting pattern supports detection without adding a visible label.
Google applies SynthID across its tools. Examples include Gemini image generation, including Nano Banana, and Veo video generation. Lyria music and NotebookLM’s podcast generation also use it. Google has described SynthID watermarking for text generated in the Gemini app and web experience.
A visible Gemini mark and an invisible SynthID watermark are separate signals. Google’s Gemini watermark guidance says AI media created or edited in Gemini Apps includes SynthID. Gemini also adds content credentials describing its origin. A file can therefore carry more than one kind of AI information.
How to check an image for SynthID
Upload the image to SynthID Detector or ask Gemini to verify it. Keep the original file so you can compare the result with the version your store plans to publish.
Use SynthID Detector
Google’s Detector access announcement makes the tool available globally in English. It covers supported content from Google and participating partners. That is broader than the Google-only SynthID check described in Gemini’s help page.
- Open SynthID Detector.
- Upload the image you want to check.
- Let the detector analyse the file.
- Read the result and any indication of which areas contain a detectable watermark.
Google’s Detector explanation describes the upload, scan and result process. Treat the result as evidence about supported watermarks. A detector cannot certify that every unmarked image came from a camera.
Check in the Gemini app
Use a signed-in personal account or an eligible Workspace account. Google’s verification instructions specify the following process.
- Open Gemini and add one image file. The file must be 100 MB or smaller.
- Ask whether Google AI created or edited the image.
- Read the verification response, including any qualification about partial detection or an unclear result.
A detected SynthID watermark means Google AI created or edited all or part of the image. An absent signal does not rule out other AI systems. Google also names very simple images and minor edits as reasons detection can be unclear.
A product-image example
Suppose you use Gemini to put a catalogue photograph of a blue chair into a generated living room. This is a hypothetical workflow, not a reported detector test.
Save the generated file, then upload it for verification. If Gemini detects SynthID, record that result beside the image’s editing history. It shows Google AI involvement somewhere in the image. It does not establish that the chair’s fabric or dimensions match the item you sell.
Compare the chair against the source photograph before approving the image. Repeat that comparison on the final store version, because resizing or later edits can change what customers see.
SynthID compared with content credentials
SynthID embeds a signal in the content, while content credentials describe its origin and editing history in a signed record. The Coalition for Content Provenance and Authenticity (C2PA) defines the standard behind those credentials.
What each signal tells you
| Question | SynthID | Content credentials |
|---|---|---|
| Where is the information? | Within the generated content | In a signed record commonly embedded in the file or linked externally |
| What does a check establish? | Whether a supported watermark is detected | Whether the signed origin and history information validates |
| Can edits affect the signal? | Extensive changes can prevent detection | Editing software can remove or fail to update embedded credentials |
| Does it prove a product claim? | No. It identifies AI involvement | No. It records provenance, which means origin and history |
The C2PA explainer explains that credentials can also be recovered through supported watermark or fingerprint systems. The distinction is between the embedded signal and the history record, rather than a rule that credentials always disappear when file metadata changes.
Stores encounter both because an image can move from an AI generator through editing software to a product feed. Keep the original download and preserve its provenance information through that process. The content credentials guide explains the file-level record in more detail.
SynthID and AI product images
An AI product image can carry SynthID, file metadata and a platform’s public AI label at the same time. Each has a separate job. Your store still needs to check the image against the actual item being sold.
Google Merchant Center requires AI metadata
Google’s AI-generated content policy permits AI-generated images in its main, additional and lifestyle image attributes. It requires metadata identifying AI generation and tells merchants to preserve embedded source-type tags.
The policy uses the International Press Telecommunications Council (IPTC) property DigitalSourceType. Its listed values distinguish generated content, synthetic composites and purely algorithmic media. For an image made with a trained generative model, Google specifies TrainedAlgorithmicMedia.
A detectable SynthID watermark does not replace that metadata requirement. Keep the required tags in the file your feed actually serves, including after image conversion or optimisation.
Google’s main-image rules also require the image to show the exact item being sold. They prohibit promotional overlays and visible graphic watermarks. The invisible SynthID signal is different from those visible stamps. When AI changes a product’s colour or adds an accessory you don’t sell with it, fix the image before submitting it.
Social platforms can add their own labels
Meta’s AI-labelling policy describes AI info labels on Facebook and Instagram. Meta uses industry signals or the uploader’s disclosure to identify AI content. Its treatment distinguishes generated content from content that was only edited with AI.
That policy does not establish that Meta directly reads every SynthID watermark. File credentials and a platform’s own disclosure controls can also affect the label customers see. Inspect the published post through the platform’s normal review process instead of assuming its label will match a Gemini result.
For your store’s own product page, add an AI disclaimer where customers need to understand a generated scene. If the room is illustrative, say so beside the image. Keep the product description tied to the real chair, including its actual dimensions and materials. Before publishing, approve the final image for product accuracy and confirm that the file retains the AI metadata its destination requires.
Give your store its own shopping agent
Customers ask in their own words. Your agent finds, compares and checks out from your catalogue, right on your site.