Three marks, three
different questions.

The short answer

Claude’s announced watermark and Google’s SynthID-Text are close relatives — the same statistical method, run by different companies with different keys. C2PA is something else entirely: signed metadata about a file’s history. The practical difference is what question each can answer. The watermarks answer “was this provider’s model involved in choosing these words?”; C2PA answers “what does this file say about how it was made?”

SynthID-Text: the published original

Google DeepMind published SynthID-Text in Nature in October 2024 and has deployed it in Gemini. During generation, a secret key and recent context seed pseudorandom scores, and a tournament nudges token selection toward candidates that score well. Across a long passage those nudges form a detectable pattern, while the text reads normally — the paper reports the method is non-distortionary on average, and Google measured no significant quality difference across roughly 20 million live responses.

Google also open-sourced a reference implementation with example detectors. Detection needs the text and the matching key, but not the model itself.

Claude’s watermark: the same method, a different key holder

In August 2026, Anthropic announced that future Claude models will carry a watermark it describes as a version of SynthID-Text, motivated by the EU AI Act’s transparency obligations. Anthropic states the mark adds no tokens, uses no hidden characters, and carries no user or organisation identity. A detection API was announced as forthcoming.

Because each deployment is keyed separately, the two systems do not interoperate: Google cannot test for Claude’s watermark and Anthropic cannot test for Gemini’s. “Watermarked” is always watermarked by someone, and only that someone can check. The mechanics are covered in depth in our technical explainer.

C2PA: provenance for files, not words

C2PA content credentials take a different approach: a cryptographically signed manifest attached to a file, recording assertions about how it was created and edited. Anyone with a C2PA-aware verifier can check a credential — no secret key custody, no statistics.

The trade-off is durability. The credential belongs to the file, not the content. Copy the words into an email and the credential stays behind; strip the metadata and it is gone. For prose that travels by copy-paste — which is most prose — C2PA rarely survives contact with reality. Its absence therefore proves nothing, and for text it complements rather than replaces generation-time watermarks.

Choosing the right check

• “Was Claude involved in this passage?” — only Anthropic’s keyed detector can say, once its API ships.
• “Was Gemini involved?” — only Google’s SynthID detector.
• “Is this the original file, and what does it record about its making?” — verify the C2PA credential, if one is attached.
• “Are there hidden characters in this text?” — none of the above; that is a simple character inspection, covered in our invisible-characters guide.