C2PA vs SynthID

These two technologies get mentioned in the same breath constantly, which is unfortunate, because they are close to opposites. Nearly every practical question about AI watermarking becomes easy once you see how they differ.

The core difference

C2PA attaches provenance to a file. A manifest is embedded alongside the picture, containing claims about origin and edits, signed by an issuer. It is a label on the bottle.

SynthID embeds provenance in the content. The pixel values themselves are adjusted in a recoverable pattern. It is dissolved in the liquid.

Everything else follows from this.

How they compare

	C2PA	SynthID
Stored in	Metadata attached to the file	The image content itself
Readable by	Anyone with a C2PA reader	Google’s detector only
Survives metadata stripping	No	Yes
Survives screenshots	No	By design, yes
Survives re-compression	Usually not	By design, yes
Human-readable content	Yes — tool names, edit history	No — just a detection signal
Can be verified independently	Yes, with trust lists	No
Removable with this site	Yes, verifiably	No, and we say so

What each is actually for

C2PA answers: what does the producer claim about this file, and can I check that claim has not been tampered with? It is a positive-assertion system. A newsroom or a photographer uses it to prove their own content’s history. Its weakness is fragility — the manifest is trivially lost, and its absence proves nothing.

SynthID answers: is this one of ours? It is a recognition system that works in the provider’s own interest, and its strength is exactly C2PA’s weakness: it survives the messy real-world journey a file takes. Its limitation is that it is opaque and closed — only the issuer can read it, and it carries no human-meaningful information.

Why this determines what tools can do

A metadata tool like this one operates on the container: the segments and chunks wrapped around the compressed picture. It can read and remove C2PA cleanly, and it can verify the removal by re-scanning.

It cannot touch SynthID, because it never touches pixels at all. That is by design — copying image data byte-for-byte is what makes cleaning lossless.

So on our scanner, C2PA gets a real status — Detected or Not detected — while SynthID permanently reads Unable to verify. Those are not different degrees of confidence. They are different kinds of statement: one about the file, one about our instruments.

The practical takeaway

If you are checking an image’s origin, C2PA is what you can actually inspect, while remembering its absence means little.

If you are cleaning an image for privacy, metadata removal genuinely helps: GPS, device identifiers, timestamps and prompts all go.

If you are hoping to make AI-generated content unidentifiable, neither this site nor any other browser tool will get you there, and you should be suspicious of anything that claims otherwise. See Can you remove AI watermarks? for why.