humanflow

Can SynthID or C2PA prove content is AI-generated?

They can provide positive evidence that a participating model or application produced something — and neither can establish the opposite, because the absence of a watermark or a credential is the normal state of almost all text.

Last reviewed 15 August 2026 · The HumanFlow team

The two work differently and fail in the same direction. SynthID marks output as a model generates it, by adjusting the probability scores assigned to candidate next words so a detectable pattern is left in the text itself. C2PA attaches signed provenance data alongside a file — Content Credentials, which its own documentation likens to "a nutrition label for digital content".

Both are therefore evidence about origin rather than inferences about style, which makes a positive result far stronger than any detection score. If a watermark is present, something that plants watermarks produced the text. If a credential validates, a named party has signed a statement about how the file was made.

The asymmetry is the whole answer. A watermark exists only where the model's operator chose to plant one, so anything from a model that does not participate — including open-weight models anyone can run locally — carries nothing. Credentials are stripped by a screenshot, a re-encode, or copying text out of one application into another, which is how most text moves.

So neither can prove a human wrote something, and neither can prove nobody used AI. For text specifically, C2PA is the weaker of the two, because provenance travels with a file and a pasted paragraph arrives bare regardless of its history.

When this answer changes

It changes if watermarking becomes near-universal across model providers, which would make its absence more informative than it is today. That is a coordination problem rather than a technical one, and open-weight models are a standing exception to it.

It does not change for the report in front of you. Neither system is what produced an AI writing score on an academic integrity report — those come from style-based detectors, which is a different technology answering a different question.

Where to go next

  • SynthIDWhat its documentation states about the mechanism, and what it conspicuously does not state.
  • C2PA and Content CredentialsWhat a credential asserts, and why its absence is the normal state of most files.
  • WatermarkingThe general technique, with a worked example of how the signal is embedded.
  • What paraphrasing does to detectionThe paper that showed the attack also showed the defence, and it is usually quoted at half length.

Sources

One of our direct answers.