Technical guide
C2PA vs SynthID: metadata and pixel watermarks explained
C2PA Content Credentials and SynthID can both carry signals about digital media, but they are not interchangeable. C2PA packages signed provenance claims in a manifest that may be stored as metadata. SynthID embeds a signal in generated content or token choices and needs a compatible detector.
C2PA vs SynthID at a glance
| Property | C2PA Content Credentials | SynthID |
|---|---|---|
| Primary purpose | Record provenance claims and edit history | Mark content produced by supported Google AI systems |
| Location | Manifest embedded in or referenced from an asset | Signal embedded in image/audio/video pixels or text token patterns |
| Human-readable information | Can expose claims through a compatible viewer | Usually produces a detector result, not a provenance history |
| Detection | Parse and validate the manifest and its bindings | Use a compatible detector profile |
| Can ordinary metadata tools find it? | They may locate embedded containers, but full trust needs validation | No; the signal is not an EXIF/XMP field |
| Covered by Watermark Tidy | Supported embedded PNG/JPEG containers and selected metadata fields | Verified reference engine; provider profiles still required |
C2PA is a provenance framework
A C2PA manifest contains assertions, a claim, a signature, and bindings between the claim and the asset. Assertions may describe capture, editing actions, tools, ingredients, or AI involvement. A manifest store can be embedded in a file or referenced externally.
Finding a C2PA container does not by itself validate it. Full verification checks structure, signatures, signing credentials, trust chains, claims, and content bindings. It also does not imply AI generation: cameras and human editing software can produce Content Credentials.
SynthID and Claude text marks need compatible detectors
Google describes SynthID as a family of watermarking and detection techniques for AI-generated images, audio, video, and text. Image marks are embedded in pixels so they can survive some transformations. SynthID-Text changes token probabilities during generation, creating a statistical pattern a compatible detector can test.
Anthropic confirms that Claude uses a version of SynthID Text and offers its official detector in private preview. The exact Claude tokenizer, keys, profile, thresholds, and test vectors are not public, so an unrelated reference score cannot become a Claude verdict. The Claude text detection guide explains the available local checks and this boundary.
None of these signals is an ordinary metadata field. Removing C2PA metadata does not remove SynthID, and deleting invisible Unicode does not verify removal of a statistical model-level text signal.
What removal can and cannot establish
- Removing embedded C2PA metadata may remove local provenance information, but external or soft-bound manifests may still exist.
- Missing Content Credentials do not prove that content is authentic, human-made, or free of AI involvement.
- A negative SynthID result applies only to the compatible detector and supported generation system.
- Metadata removal and pixel-watermark removal are different technical operations.
Watermark Tidy handles only supported hidden Unicode and selected PNG/JPEG metadata structures. It does not attempt pixel-level or statistical removal.
Sources: C2PA Content Credentials specification, Google DeepMind SynthID, and Anthropic's guide to Claude content marks. Last reviewed September 3, 2026.