The pattern is created while text is generated.
1Many plausible next tokens
The model has several reasonable word choices.
2Consistent hidden preference
A watermarking rule slightly favors one set.
3Detector tests a passage
A matching detector looks for aggregate evidence.
Text patterns and file credentials solve different problems.
Statistical text watermarking puts a signal into the generated language. C2PA instead attaches a signed manifest to a file. The manifest can name a claim generator, record actions, and carry signature information.
| Signal | Lives in | Openly inspectable? |
|---|---|---|
| Statistical watermark | Word/token pattern | Only with a matching detector |
| Hidden Unicode | Text characters | Yes |
| C2PA credential | Signed file manifest | Yes, if preserved |
Editing changes the evidence, but not in a binary way.
Short text provides less statistical evidence. Translation, heavy rewriting, mixed authorship, and constrained factual text can also affect detection. File credentials may be stripped by export or upload. These limits are why a result should not be treated as proof of identity, intent, academic misconduct, or factual truth.