09/22/2026
C2PA published an updated implementation guide for Content Credentials on August 11, closing gaps that made the previous version easy to satisfy on paper and hard to rely on in practice. Five changes matter: a shared IPTC-based classification for how content originated, a richer ai-disclosure assertion covering model provenance and human oversight, region-level localization of AI edits inside a single asset, verifiable generation-recipe metadata, and formal vocabulary for created/opened/edited actions.
OpenAI documents SynthID watermarking plus C2PA metadata for images and audio, with text support still a stated goal rather than a shipped feature. Anthropic has gone further, shipping an imperceptible text watermark in production alongside signed C2PA metadata for images, tied explicitly to its Article 50(2) Code of Practice commitments.
None of this is a compliance shortcut. Article 50 of the EU AI Act, in force since August 2, splits obligations between providers, who must mark synthetic output, and deployers, who must disclose AI involvement to their own readers when text on a matter of public interest skips human review. A vendor's watermark answers the first question. It does not answer the second. If AI provenance sits in your governance framework, ask vendors which of the five assertions in the August guide they actually implement, whether their signal survives your publishing pipeline, and separately, whether your organization is making the disclosure Article 50 requires of you as a deployer, watermark or no watermark.
