Anthropic to start watermarking Claude-generated text, images

Anthropic PBC has announced plans to embed invisible watermarks in text and images generated by its AI model Claude. This move aims to bring Claude into compliance with the European Union's AI Act, which includes a voluntary Code of Practice that Anthropic has signed. The watermarking mechanism will be introduced in Claude models released after August 2 and will also extend to earlier models in the future.

The watermarking process is designed to be invisible and will persist through copying, pasting, and even editing of the text. While the exact method of text watermarking remains undisclosed, Anthropic may be utilizing a similar technology to Google DeepMind's "SynthID." This technology modifies a language model's word choices to make the generated text detectable. SynthID supports not only text but also AI-generated images, videos, and audio, and is integrated into several Google consumer AI services.

For media files, Anthropic plans to use the Content Technology Platform (C2PA) developed by Google. C2PA marks AI-generated images with metadata files specifying the model that created them, the creation time, and any copyright restrictions. Developers can add additional details to the metadata. C2PA incorporates mechanisms to prevent hackers from tampering with the metadata, such as creating a unique identifier (hash) and detecting changes to the entire metadata file.

Anthropic intends to release tools to help users detect Claude-generated content, although the company warns that the technology may not always detect watermarks in short or heavily edited text snippets. OpenAI Group PBC, another major player in AI, has also signed the AI Act's voluntary Code of Practice. OpenAI may launch its own text watermarking mechanism soon, as it created a similar feature in 2024 and participates in the consortium that develops the C2PA standard.