What is SynthID?

The idea

Most provenance information is attached to a file. C2PA content credentials, EXIF data, XMP — all of it lives in a metadata container wrapped around the actual content, which makes it easy to read and equally easy to destroy. Screenshot the image and every bit of it is gone.

SynthID takes the opposite approach. It modifies the content itself in ways below the threshold of human perception: minute adjustments to pixel values in an image, to the waveform in audio, or to word choice in text. There is nothing to strip, because the mark is the content.

What it covers

• Images — from Google's image models, and since 2026 from OpenAI's as well.
• Audio — embedded in the waveform, surviving compression and re-recording to a degree.
• Video — applied frame by frame.
• Text — the same statistical token-choice approach Anthropic uses for Claude.

How durable is it, really?

More durable than metadata, which is a low bar, but not indestructible. Google's own description is that it survives “many common modifications” — cropping, resizing, compression, colour adjustment, adding filters. That is genuinely impressive, and it is exactly the class of operation that annihilates a C2PA credential.

It is not magic, though. Aggressive enough transformation degrades the signal, and for text, heavy paraphrasing defeats it, since rewriting the words removes the very thing carrying the mark.

Who can actually detect it

This is the practical limitation, and it is the reason SynthID appears on our signals list as something we cannot read.

Google runs a SynthID Detector portal, announced in May 2026, where you can upload a file and see which parts of it carry the watermark. But it began as an early-tester rollout with a waitlist prioritising journalists, media professionals, and researchers, and Google has not published an API for third-party developers to build on.

So the honest position is not “no detector exists” — one does, and it is Google's. It is that you cannot check SynthID from a general-purpose tool, and neither can we, because the capability is not offered to anyone outside that gate.

A confusion worth clearing up

Google has open-sourced the SynthID text watermarking method, which sounds like it should solve the detection problem. It does not. Open-sourcing lets other developers build watermarking into their own models — it is a tool for marking your own output, not a key to reading Google's. Detecting a watermark generally requires the parameters used to embed it.

The limitation everyone forgets

SynthID only ever finds content from models that use SynthID. That is mostly Google's tools, plus OpenAI's images. It knows nothing about output from Meta's models, Midjourney, Stable Diffusion, or any of the countless open-weight models running on someone's own hardware.

Which produces the same trap as every other signal in this field: a negative result is close to meaningless. “No SynthID detected” means “not made by a model that uses SynthID,” and that is an enormous, ever-changing set of everything else — including every image ever taken by a camera.

What to do instead, today

Check what is actually checkable. If a file still has its C2PA credential, that is cryptographic evidence of origin and integrity, and it is readable right now in your browser. Metadata gives you context. A file hash gives you identity. None of it tells you whether AI was involved unless the producer chose to declare it — and that declaration, when present, is worth more than any detector's guess.

When Google opens a public API, we will add SynthID to the audit and say so on the signals list. Until then it stays marked as unreadable, because that is what it is.