The short answer: metadata and watermarks are different layers
C2PA Content Credentials carry a signed claim: who issued the assertion, which actions were declared, which ingredients were used and whether technical validation succeeded. The manifest lives in the media container and may be lost through re-export, screenshots or metadata cleanup.
SynthID works differently. Google DeepMind describes it as an invisible digital watermark embedded directly into an image, audio, text or video. In an image, the signal is part of the media rather than an EXIF/XMP field, so deleting metadata sections does not target it.
What C2PA can say — and what its signature cannot prove
C2PA is useful because it can carry rich context. A validator can check the claim signature and its binding to a particular asset, while an interface may show the application, declared actions and ingredient chain. A technically valid signature does not establish that the depicted event is true or turn the claim generator into an independent fact-checker.
A supported manifest can be removed from a separate derivative without rewriting its contents. The result is a file with no accessible embedded claim, not a new history saying “made by a human.” An honest replacement chain requires another signer to issue a new claim and connect it to the source.
Why an EXIF cleaner does not remove SynthID
An EXIF cleaner edits structured container fields. SynthID is designed to remain detectable through some common transformations such as cropping, filters or lossy compression. That does not make a watermark indestructible: OpenAI’s official material explicitly allows for signal degradation, and a negative detector result does not exclude generation by a supported system.
ClearTags currently has no SynthID detector and must not display a “watermark absent” verdict. It can only show accessible C2PA, EXIF, XMP, IPTC and container data. Treating clean metadata findings as a SynthID check would be a false inference.
Why one file can carry both signals
The mechanisms complement each other. C2PA carries detailed, signed, machine-readable context, while a watermark can preserve a narrower provenance signal when container metadata is lost. OpenAI documents using C2PA and SynthID together in images generated by ChatGPT, Codex and its API.
Detection is still implementation- and source-specific. OpenAI verification looks for signals associated with OpenAI, while Gemini describes detecting SynthID in Google AI media. Neither statement describes a universal detector for every image from every generator.
FAQ: what should you check before publishing?
Start with the goal. For privacy, remove GPS, creator identity, prompts, workflows and supported C2PA from a separate publishing copy. For process evidence, preserve the original and validation report. For platform compliance, classify the actual content and use required disclosure instead of relying on removal of one technical signal.
- Can C2PA be removed? A supported embedded manifest can be removed from a new copy; it cannot be silently rewritten.
- Can an EXIF cleaner remove SynthID? No. It is not an EXIF/XMP field, and ClearTags does not offer that operation.
- Does an empty report prove no AI was used? No. A signal may be absent, lost, degraded or unsupported.