Claude will apply invisible watermarks to AI text and images

Artificial intelligence developer Anthropic has announced plans to integrate invisible, machine-readable identifiers into both the text and image outputs produced by its Claude models.

This strategic initiative is designed to align the company's products with the stringent transparency regulations established by the European Union under its newly enacted AI Act. While these digital markings will remain completely imperceptible to human users, they will allow software applications and online platforms to easily determine whether a piece of media or written copy was generated by an artificial intelligence. The organization plans to implement these safety features automatically in all future model releases, while progressively working to retrofit its existing suite of tools during the regulatory grace period.\n\nTo mark visual media, the technology company will adopt the established C2PA metadata standard, a provenance framework that is already backed by several major industry players. For textual outputs, Anthropic intends to weave an imperceptible digital signature directly into the generated text without altering its fundamental meaning, quality, or readability. This signature is designed to travel with the text even when it is copied, pasted, or slightly edited. Despite these technological advancements, the ultimate effectiveness of such watermarking techniques remains highly debated. Industry experts note that metadata can easily be stripped from media files, and even Anthropic acknowledges that these detection methods are not entirely foolproof, meaning unmarked content could still easily originate from generative models.