C2PA vs SynthID: two checks that should stay separate
Understand the difference between inspecting a signed manifest and detecting a provider watermark.
The short answer
C2PA attaches signed provenance records; Google DeepMind's SynthID embeds watermarks in supported AI-generated content. They are different mechanisms and can coexist. C2PA Studio inspects embedded C2PA records; it does not detect SynthID. An absent C2PA manifest cannot establish whether a file is AI-generated or watermarked.
Source context: Google DeepMind: SynthID. The workflow recommendations below are editorial analysis.
One file can have more than one evidence channel
C2PA records and invisible watermarks are often placed in the same conversation, which makes a single “AI detector” button sound convenient. In practice, you should name the mechanism you actually tested. A manifest reader examines embedded structured credentials. A watermark detector looks for a particular signal using its own technology. Results from one should not be relabeled as results from the other.
Google describes SynthID as a watermarking system for generated media. This inspector does not implement Google’s detector. If the manifest mentions a generator, that is a declaration you can read; it is not a successful detection of a watermark in the image, video frames or audio.
Design a two-column review
For a team evaluating a generated clip, create two separate evidence entries. The first records embedded credential presence, validation messages and the tool used to read them. The second records whether an authorized provider detector was used, its result and the date. If a check was not performed, write “not tested” rather than converting silence into a negative.
Keep each result tied to the same exact file. A detector run on the original and a manifest check on a platform download are not directly comparable. Save both copies if you need to test a transformation, and repeat each applicable check after the transformation. This removes an easy source of contradictions in production reviews.
What disagreement tells you
Suppose a provider detector identifies a signal while the local inspector finds no embedded credentials. That combination is not inherently contradictory. You have tested different mechanisms. Your next action is to document the distinction and check whether the download path carried the source metadata. Do not promise to restore an original signing history from the detector’s output.
For procurement, ask a vendor a specific question: what can be inspected on the exact export format and endpoint we will use? Request a sample and a documented verification route. Keep marketing descriptions separate from the evidence your team reproduces. This is a stronger acceptance criterion than asking whether the vendor generically “supports AI transparency.”
A comparison for your review
| Task | Manifest inspection | Provider watermark check |
|---|---|---|
| Read declared workflow actions | Inspect available assertions | A separate mechanism |
| Detect SynthID signal | Not implemented here | Use a documented provider route |
| Check an export | Repeat on the delivered copy | Repeat where supported |
Put it into practice
- Name each detection mechanism explicitly.
- Use the exact same file for comparable checks.
- Mark checks not performed as not tested.
- Keep provider claims separate from your reproduced observations.
Bring your own evidence.
Inspect a local file and see what it actually declares.