What is SynthID, and how does it actually work?

The short answer

SynthID is a watermark made by Google DeepMind. It is placed inside the content itself. In an image it goes into the pixels. In audio it goes into the sound. In video it goes into the frames. In text it goes into the word choices.

You cannot see it or hear it. But a matching detector can find it.

How it is different from the mark you can see

	Visible sparkle	SynthID
Where it sits	One corner	All over the file
Can you see it	Yes	No, by design
Survives a crop	No, just cut it off	Mostly yes
Survives saving again	Yes	Mostly yes
Can this site remove it	Yes	No

Why a spread-out watermark is so hard to remove

A logo is a local change. A few thousand pixels in one place change, and the rest of the picture stays the same. Find that place, undo the change, done.

SynthID is not local. It changes a very large number of pixels by a very small amount. Each change is far too small for your eye to notice. It is even smaller than the noise that JPEG compression adds anyway. But the changes follow a pattern, and the detector knows that pattern.

So there is no box to point at. There is no area to rebuild. That is also why normal edits do not kill it. Saving as JPEG, resizing, screenshots, colour changes, filters — it is built to survive all of that. Push hard enough and it will break. But by then your picture is usually ruined too.

Where it is used

• Images from Google’s image models, including what the Gemini app makes.
• Video from Veo, carried across the frames.
• Audio, placed in a part of the sound you do not hear.
• Text, by nudging the choice between words that are equally good. This works better on long text than on one line.

What SynthID does not tell you

Most explanations stop before this part. It matters.

• It is not a general AI detector. SynthID finds SynthID. An image from another AI tool comes back clean. Clean means “no Google watermark found”. It does not mean “made by a human”.
• Nothing found proves nothing. Heavy editing or a bad re-save can leave an AI image looking unmarked.
• It does not say who or when. For that you need C2PA content credentials. Those are a signed record, not a hidden signal. They can be stripped. SynthID cannot.
• It is not about rights. It carries no licence or usage information.

How to check an image for it

You cannot check it yourself with an EXIF viewer. There is nothing written in the file to read. Finding it needs the matching model, so it happens through Google’s own tools. You can ask about an image inside the Gemini app. Google also runs a SynthID detector portal for journalists and researchers.

Step by step here: how to check if an image is AI-generated.

So what does this mean for watermark removal?

It keeps it honest. Taking the sparkle off is a visual fix. You get back a clean corner and the pixels that were under it. That is all. It does not hide that the image is AI-made. Any browser tool that claims to strip SynthID is talking about the logo and calling it something else.

Questions people also ask

Can SynthID be removed?

Not by any browser tool, and not by the usual tricks. Screenshots, format changes, resizing and filters are exactly what it is built to survive. Enough damage will break it, but your picture will be ruined first.

Does removing the Gemini watermark remove SynthID?

No. They are two separate layers. Rebuilding the pixels under a corner logo leaves the hidden signal in the rest of the picture completely untouched.

Is SynthID in every Gemini image?

Google says it marks the output of its image, video and audio models. So treat anything from those tools as carrying it. And do not treat its absence elsewhere as proof of anything.

Can SynthID tell me which model made an image?

It tells a detector that a Google model marked it. For who, when and with what, you need C2PA content credentials — if they are still attached.