What is SynthID? Invisible watermarks explained
SynthID hides a signal inside the pixels or audio of AI-generated media. See how invisible watermarks differ from metadata and what they mean for creators.
Metadata sits beside the picture. A watermark like SynthID sits inside it. That difference explains why some AI content is still recognised after its file information is gone.
What SynthID is
SynthID is a watermarking system from Google DeepMind, first announced in 2023 for images made with Google’s Imagen model. It has since been extended to video, audio and text produced by Google’s AI tools. As the model generates the content, it makes tiny adjustments that people cannot see or hear but that a detector can recognise as a pattern.
How it differs from metadata
| Metadata (C2PA, IPTC) | Invisible watermark | |
|---|---|---|
| Where it lives | Attached to the file | Inside the pixels or sound |
| Who can read it | Anyone with a reader | Those with the detector |
| Survives a screenshot | No | Usually |
| Carries detail | Yes: tool, edits, author | Very little: mainly “made by this system” |
| Works across companies | Yes, open standard | Mostly per provider |
The two are meant to work together. Credentials explain a file’s history when they are present. A watermark remains when they are not.
Who checks for it
Google provides detection for content made with its own models, including a verification feature in its Gemini app and a detector portal for media organisations and researchers. Other companies are building similar systems. In 2026 TikTok said it had started adding its own invisible watermarks to AI content made with its tools and to uploads that arrive with Content Credentials, so that the label can follow a video after it is downloaded and reposted.
What this means for creators
- A clean metadata check is not the whole story. Our checker can tell you a file has no AI markers in its metadata. It cannot tell you whether the pixels carry a watermark.
- Labels are getting stickier. As platforms adopt watermark detection, AI content will be recognised even when its file data has been lost in sharing.
- Real photos are not affected. A watermark is only present if a watermarking model produced the content. A camera photo with conventional edits has nothing to detect.
- Mixed edits can carry it. If part of an image was produced by a watermarking model, that part can be detected.
Is it reliable?
Watermarks are a strong signal when found: a positive result is good evidence that a specific system produced the content. The reverse is weaker. No watermark does not prove something is real, because many generators do not watermark and heavy processing can degrade the signal. That is why platforms combine watermarks, Content Credentials, their own detection and creator disclosure.
For creators the practical lesson is simple. Disclosure is becoming something the content carries with it. Labelling AI work yourself is easier than being caught out by a system you cannot see. See how to label AI content.
Questions people ask
Can I see a SynthID watermark?
No. It is designed to be imperceptible. It can only be detected with the matching detection tool.
Can this site detect SynthID?
No. Our checker reads metadata. Invisible watermarks live in the pixels or audio and need the provider’s own detector.
Does every AI image have SynthID?
No. SynthID is Google’s system and is applied to content from Google’s models. Other companies use their own watermarks or none.
Does saving or cropping an image remove it?
The watermark is built to survive common changes such as compression, resizing, cropping and filters, though no watermark is indestructible.
Is SynthID the same as Content Credentials?
No. Content Credentials are attached data that can be read by anyone and can be lost. SynthID is part of the content itself.