AI Provenance Signals: SynthID, C2PA, and the Limits of Watermark-Based Detection

What C2PA (Content Credentials) is

C2PA is an open technical standard, developed by the Coalition for Content Provenance and Authenticity, for attaching a cryptographically signed record of a file's origin and edit history. When present and unmodified, a Content Credential can show what device or tool created a file and what edits were applied.

What SynthID is

SynthID, developed by Google DeepMind, embeds an imperceptible digital watermark directly into pixels generated by participating models, designed to remain detectable after common transformations like cropping or recompression.

What these signals can't do

Both signals depend entirely on participation: C2PA requires the capturing or editing tool to sign the file; SynthID requires the generating model to embed the watermark. Neither signal is retroactively added to content from a non-participating tool, and both can be stripped by platforms that don't preserve metadata during processing. Absence of either signal is common and uninformative on its own — it is not evidence of manipulation or AI generation.

• Only meaningful when the originating tool participates in the standard
• Can be stripped by upload pipelines that don't preserve embedded data
• Prove a claim was made and signed — not that the claim is true

Why general-purpose detection still matters

Because most images in circulation — from non-participating cameras, older devices, non-participating generators, or platforms that strip metadata — carry neither signal, provenance standards and general-purpose AI detection are complementary, not competing: provenance proves origin when present; detection estimates likelihood when it isn't.

Does a missing C2PA credential mean an image is fake?

No. Most images never had one attached, or had it stripped during upload — a missing credential is common and uninformative on its own.

Can SynthID detect images from any AI generator?

No. It can only identify content from models that embed the watermark at generation time. An image from a non-participating generator will show no SynthID signal regardless of how it was made.

Does PhotoProof AI check for these signals?

PhotoProof AI's metadata review layer inspects available provenance and embedded metadata as one of five evidence signals — see the multi-signal decision making page — but does not depend on any single provenance standard being present.