Anthropic adds invisible watermarks to Claude text and images

Anthropic is rolling out a system to embed invisible identification markers into content created by its Claude AI models, bringing the company into line with new European regulatory standards that demand AI transparency.

Meeting EU requirements#

The European Union's AI Act introduced fresh labeling and disclosure rules effective August 2nd, compelling AI makers to signal when content has been artificially generated. While existing Claude models have a four-month grace period to comply, all newly released versions will apply these marks from launch. The shift reflects growing regulatory pressure on AI vendors to make their tools' output verifiable and traceable.

How the marking works#

Anthropic is deploying two complementary approaches. For images, the company adopts C2PA, an established provenance standard already used by Adobe, OpenAI, and Google. For text, Anthropic embeds what it calls an "imperceptible watermark" directly into the generated output—a technique that preserves the content's meaning, quality, and readability while remaining invisible to human readers.

According to the company, these text watermarks will remain attached to Claude's output across all product surfaces, including its API platform and versions accessed via AWS, Google Cloud, and Microsoft Foundry. Because the markers are embedded in the text itself, they should theoretically persist even when content is copied, pasted, or lightly edited elsewhere online.

Detection challenges ahead#

Anthropic promises to release technical documentation on how third parties can identify both watermarks and C2PA metadata, though the company hasn't yet confirmed whether existing detection tools—such as those built into Google's Gemini—will recognize Claude's markers.

The watermarking initiative faces real-world obstacles. C2PA metadata can be stripped away, sometimes unintentionally, when files are uploaded to platforms. Anthropic itself acknowledges that its system is imperfect: content lacking detectable marks may still originate from generative AI, meaning the absence of a watermark proves nothing. The effectiveness of these measures will ultimately depend on their resilience across diverse digital platforms and workflows.

This development aligns with broader industry momentum toward AI origin labeling. Community-led projects, such as detector tools emerging on fanfiction platforms, have already begun experimenting with similar identification systems—though they operate at a much smaller scale.