What can and cannot be verified
Four signals. Different evidence. No shortcut from a character count to a provenance claim.
| Signal | Evidence / trust | Watermark itself? |
|---|---|---|
| Invisible Unicode character operations | High: deterministic and independently repeatable. | No. An enumeration of actual characters. |
| File C2PA Content Credential | High: cryptographic validation of a supported credential. | A file-side mark, not the text watermark. |
| Own-key residual measurement | Medium: reproducible on our controlled bench, not official. | An indirect measurement of our own watermark signal only. |
| Third-party classifier “AI probability” | Low for watermark claims: a classifier's guess. | No. It does not read the watermark. |
Only one of these four is a measurement of the watermark itself, and it is not the one that comes with a green checkmark.
Why third-party detector scores are not reliable evidence of a watermark
A classifier's guess and reading a watermark signal are different operations. This site supplies no AI probability, classifier verdict, or confidence percentage.
Lack of a detected mark doesn't mean the content wasn't AI-generated.
The official documentation describes text being “heavily edited,” “paraphrased,” “translated,” or “mixed into other writing,” short samples “leaving too little text for a reliable signal,” stripped file metadata, and unsupported models, features or file types as legitimate reasons a mark may not be detected.
The official boundary
Claude “weaves an imperceptible watermark directly into the text itself”; it “doesn't change the meaning, quality, or readability.” It travels with copied text and may partly withstand editing. Text marking is SynthID-Text-style statistical marking. A viewable, verifiable file C2PA Content Credential is a separate mechanism.
The detection API is in private preview for eligible organizations, not a publicly accessible service on this site.
Read Anthropic's official documentation ↗ · SynthID-Text reference implementation ↗
Run a controlled own-key experiment
Your prompt becomes context for a newly generated SynthID-Text sample on our server. This does not insert a mark into the original words you submitted. We rewrite that new sample, then report its costs and own-key signal measurements. Never interpret the result as a test of the original draft's Claude provenance.
No experiment has been run.
Actual run measurements
Measured on our own bench, with our own key. This is not Anthropic's detector and cannot predict its result.
This run changed too much. Try Light mode.