A statistical signal embedded in content at production so its artificial origin can be detected, distinct from provenance, which rests on signed metadata.
Two families of technique answer the same question without being the same thing. Watermarking inserts into the content itself an invisible statistical signal a detector can recover; provenance attaches to the file cryptographically signed metadata describing how it was made and subsequently modified. The second approach is conceptually more robust since it rests on a signature, but it is fragile in practice, a screenshot or a recompression destroying the metadata, whereas a watermark survives such handling and yields to deliberate attack. Neither therefore solves the problem alone, and combining them is the only serious path. The European regulation imposes a machine-readable marking obligation on providers of systems generating synthetic content, and a disclosure obligation on deployers for deep fakes. The limit is structural and deserves stating: these mechanisms prove that a piece of content was generated, they never prove that unmarked content is authentic.
The Coalition for Content Provenance and Authenticity, formed in February 2021 by a group of technology and media companies, published a signed metadata specification since adopted by several camera manufacturers and by distribution platforms.
watermarking, filigrane numérique, marquage des contenus générés, détection de contenu synthétique