SynthID Checker: What It Is and How to Check

SynthID is Google DeepMind’s invisible watermark embedded in images, video, audio and text made with Google AI. Only Google’s detector can find it. Here you get the steps to check a file in Gemini and a free scan of what is visible without Google’s keys: C2PA and IPTC metadata and the visible mark.

To be clear: this tool does not detect SynthID. The invisible watermark cannot be read without Google’s detector — not by us, nor by any other site that has not been given access to it. We check the open signals that Google and other generators write into metadata.

How to check SynthID in Gemini

Open the Gemini app or gemini.google.com and sign in with a Google account.
Attach the file to a message. Images are supported; Google has been adding video and audio gradually, so check whether Gemini accepts your file type.
Ask directly: “Was this created or edited with Google AI?”
Gemini will say whether SynthID was found. The answer covers Google tools only: images from Midjourney, ChatGPT and other services are not recognised by this check.
For journalists and researchers Google opened the SynthID Detector portal with access on request. If you have access, upload files there.

Google changes the interface and the list of supported formats; see the Gemini Help Center for current details.

What you can and cannot see

Signal	In this tool	Where to check
C2PA manifest: signer, actions	Yes, read; signature not verified	contentcredentials.org/verify
IPTC / XMP DigitalSourceType	Yes	—
EXIF Software / Make	Yes	—
Visible Gemini icon in the corner	We show the corner, you decide	By eye
SynthID in an image	No	Gemini app, SynthID Detector
SynthID in video and audio	No	Gemini, if your file type is supported
SynthID in text	No	No public checker
Invisible watermarks of other vendors	No	Those vendors’ own tools

SynthID: What It Is and How to Check for Google AI Content

SynthID is Google DeepMind technology that embeds an invisible watermark in images, video, audio and text made by Google models. It is designed to survive compression and cropping, but only Google’s detector can read it. This page explains how to check a file in Gemini and gives you a free scan of what is available without Google’s keys: C2PA and IPTC metadata, mentions of Gemini, Imagen and Veo, and an enlarged image corner for spotting the visible mark.

Google signals in C2PA: signer, app, actions with trainedAlgorithmicMedia

IPTC/XMP DigitalSourceType, Credit and CreatorTool

Mentions of Gemini, Imagen, Veo, Nano Banana and SynthID in metadata

Enlarged bottom-right corner of the image to look for the visible icon

Step-by-step SynthID check in the Gemini app

A table of what can and cannot be checked without Google’s detector

How to Use

Upload an image or video
Review the Google signals table and the enlarged image corner
To check SynthID itself, upload the file to Gemini and ask whether it was made with Google AI

FAQ

No. Only Google holds the keys and the detector. Third-party sites promising a “SynthID detector” at best read metadata or score the image with a generic classifier that knows nothing about SynthID.

Google designed the mark to withstand common edits such as compression, cropping and filters. There is no guarantee: heavy processing can weaken it. C2PA and IPTC metadata, by contrast, disappear after a simple re-save.

No. SynthID is applied by Google models and a few partners. Midjourney, Stable Diffusion and most other generators do not use it, so a SynthID check will not flag their output.

C2PA is a signed record in the metadata: easy to read, but also easy to remove. SynthID is a signal in the content itself: hard to remove, but only Google can read it. Google uses both.

What SynthID is and how it works

SynthID is Google DeepMind’s technology for invisibly marking content made with Google AI. The watermark is embedded into the content itself at generation time, not into the metadata. People cannot see or hear it, but Google’s detector can tell from the file whether the mark is there.

Images and video. The mark is spread across pixels and frames. It was designed to survive common edits: compression, cropping, resizing and colour filters.
Audio. The signal is woven into the sound and is meant to withstand MP3 compression, added noise and speed changes.
Text. During generation the model slightly shifts its word choices according to a secret key. No extra layer is added on top of the text — the token-choice statistics are the mark. It weakens with heavy rewriting and translation, and short texts are hard to check.

The key consequence for anyone trying to check it: SynthID cannot be “read” like a metadata field. You need a detector with a key that Google holds. Any site without access to Google’s tools can only judge a file indirectly; it cannot check SynthID.

How to check SynthID: every method that actually works

SynthID in the Gemini app

The most accessible route is to upload the file to a Gemini chat and ask whether it was made with Google AI. Gemini checks for SynthID and answers in plain text. Keep the limits in mind:

the check only covers SynthID, which means content from Google tools and partners that use the technology;
Google has been extending supported file types gradually, so first see whether Gemini accepts your format;
“no watermark found” does not mean “made by a person”: another generator may have produced the file, or it may have been processed so heavily that the mark weakened.

The SynthID Detector portal

For journalists, researchers and verification organisations Google opened a separate SynthID Detector portal. Access is granted on request; there is no public “upload and check” form for everyone. If you work in a newsroom or fact-checking team, look up the current terms on the Google DeepMind website. According to Google, the portal also highlights which parts of a file carry the mark most strongly, which helps when you are looking for an inserted fragment.

SynthID in text

Google released SynthID Text as an open library for developers, so it can be added to your own language model. The detector, however, only works with the keys of the model that applied the mark. Checking from the outside whether an arbitrary text was written in Gemini is not possible. For text you are left with indirect methods such as an AI detector, which also gives a probability, not proof.

Why third-party “SynthID detectors” don’t work

Search for “synthid detector” and you will find sites promising to check any file in seconds. Without Google’s keys they can do one of three things:

read the metadata — what our tool does, just with a louder name;
run the picture through a generic “AI or not” classifier that errs both ways and knows nothing about SynthID;
return a random or always-identical answer.

If a service reports a “SynthID probability” with no reference to Google, treat it as a guess: an honest service says “we check metadata”, not “we detect SynthID”. And do not upload confidential material to it.

What to check without Gemini: indirect signals

When you cannot use Google’s check, collect what is openly visible, in this order:

Step	What to look at	What it gives you
1	The C2PA manifest, in the C2PA checker	Signer, software, the `c2pa.created` action and the source type
2	IPTC/XMP `DigitalSourceType`	A `trainedAlgorithmicMedia` label, if the file was not re-saved
3	The bottom-right corner of the image	The visible Gemini icon, unless it was cropped
4	Where the file came from	Its first appearance via reverse image search, and the author’s context

A Google signature in C2PA does not by itself mean generation: Pixel phones sign ordinary photos too, and then the source type says “captured by a camera”. The deciding field is digitalSourceType in the actions. For video and sound, the same metadata can be checked in the video watermark detector and the audio watermark detector.

The visible Gemini mark and why you can’t rely on it

Besides SynthID, some images from the Gemini app get a visible semi-transparent sparkle icon in the bottom-right corner. Our tool shows that corner enlarged so it is easy to inspect, but it does not recognise the icon automatically — you make the call. The mark cannot be relied on in either direction:

a simple crop removes it, while SynthID stays in the image;
the icon is not added in every Google product and plan, so an image obtained, for example, through the API may not have it;
a similar sparkle is easy to paint onto any picture, so its presence proves nothing either.

How to read Gemini’s answer and what to do next

Gemini replies in plain language and the wording changes between versions. Go by the meaning of the answer:

Answer	What to do
SynthID detected	The file was created or edited with a Google tool. If someone presents it as footage from a real event, that needs looking into.
The mark is only in part of the file (for example, certain segments of a video or audio track)	Likely an edit: a Google AI fragment was inserted into a real recording. Compare with the original if you can get it.
SynthID not detected	Check C2PA and IPTC metadata, run a reverse image search and look at the publication context. Other generators are not caught by this check.

SynthID vs C2PA

The two are complementary rather than competing, and Google uses both.

	SynthID	C2PA / Content Credentials
Where it lives	In pixels, sound, word choice	In the file’s metadata
Who can read it	Only Google’s detector	Anyone; the standard is open
What it tells you	Whether a Google AI mark is present	Who created and edited the file, with what and when
Screenshot, re-save	Usually survives	Usually lost
Forgery	Hard; requires the key	Signature fails without the certificate, but the record can be wiped

Should website owners worry about SynthID?

If you publish images from Gemini or Imagen, the mark is already in them and ordinary processing will not remove it. Hiding the origin makes little sense: Google does not penalise pages simply for using AI-made images; it evaluates how useful the content is. It is wiser to label such images according to the rules of the platforms where you post them and not to pass generated pictures off as documentary photos. For other traces generators leave in files and text, see the article on hidden watermarks.