Llama · Meta · C2PA
The weights are not the watermark. Meta’s key stays with Meta.
Meta signed the Code of Practice. Instagram is adding Content Credentials. That is a file. Llama is also a set of weights you can run yourself. A local completion does not arrive pre-stamped. A reply from a Meta-hosted product might.
Reviewed by Can Balkaya, CTO · Published
or type below, or drop a file
Inspected in your browser. A plan is required to finish.
Choose the surface, then the tool
| Surface | Gifi action | What can be verified |
|---|---|---|
| Ordinary words | Rewrite the prose | The words changed; no private detector certificate |
| C2PA, EXIF, XMP, properties | Strip named metadata | Action list plus an independent reinspection |
| Pixels, audio, or video | Not supported | No removal claim |
A download is not a signature
The weights do not include Meta’s detector key. If you sample Llama on your own machine, those tokens are yours unless you added a mark. Hosted Meta serving the EU is under Article 50. That output may be marked. Meta has not published the text method.
The public part is a note in the file
Instagram Content Credentials are a signed note. We can strip the local label. Soft binding can survive. A statistical trail in hosted text, if it exists, needs a rewrite.
What this cannot promise
- A rewrite does not prove the text was never AI-assisted.
- No one outside the lab can certify their detector. The key is private.
- Pixel watermarks, audio and video are out of scope.
Questions this usually raises
- Does every Llama reply have a watermark?
- No. Local weights do not. Hosted Meta might, and they have not said how.
- What about Instagram or Meta AI images?
- Those can carry C2PA. Pixels are out of scope.
- Can I delete a Llama watermark like EXIF?
- Only if it sits in the file. Word choice has nothing to delete.
Primary sources
Llama
Statistical signal? Rewrite the prose.
A Llama mark, if it is in the words, needs different words. Hidden characters and C2PA need the file cleaner.