Watermarking moves provenance into generation

Anthropic presented an invisible text-watermark approach alongside provenance metadata, bringing authenticity questions closer to the generation layer.

• The approach embeds a machine-detectable signal in generated text without displaying a visible mark to the reader.
• The release pairs watermarking with C2PA-style provenance metadata rather than treating one mechanism as sufficient.
• Detection limits and transformations remain important, so provenance is evidence—not a perfect verdict.

Provenance works best as layered evidence: generation metadata, content credentials, detector confidence, edit history, and publisher disclosure. Each layer should be independently inspectable and should fail without turning uncertainty into a false accusation.

Attach a simple provenance note to a finished piece: source material, generated components, human edits, and the credentials that survived export. Make disclosure part of authorship rather than a defensive footnote.