Anthropic to start watermarking Claude-generated text, images
Anthropic PBC has announced plans to embed an invisible watermark in text and images generated by Claude.
The Register reported the change today, citing a help desk article published on Monday. It applies to the Claude artificial intelligence model series and the Anthropic services that it powers.
The watermarking mechanism is designed to bring Claude into compliance with the European Union’s AI Act. The law, which went into effect in 2024, includes a voluntary clause known as the Code of Practice that Anthropic has signed. The provision requires model providers to mark AI-generated content.
Anthropic says text watermarking will be performed by Claude models released after Aug. 2. It plans to bring the capability to earlier models further down the line. According to the company, the watermark is invisible and persists if the text that contains it is copied to other applications. It may even remain in place if users make edits.
The company didn’t specify how the text watermark works. However, a research paper released by Google DeepMind in 2024 may provide clues. The paper describes a technology that modifies a language model’s word choices to make the text it generates detectable. Anthropic’s watermarking mechanism may use a similar method.
DeepMind’s watermarking mechanism supports not only text but also AI-generated images, videos and audio. The technology, which is known as SynthID, is integrated into several of Google LLC’s consumer AI services. Anthropic likewise plans to mark media files generated by Claude. The company’s Monday help desk article lists JPG, PNG and SVG among the image formats covered by the initiative.
Anthropic will mark images using a technology called C2PA. It works by pairing AI-generated images with a metadata file that specifies what model created them, when and whether any copyright restrictions apply. Developers can also add in other details.
C2PA includes multiple mechanisms designed to prevent hackers from tampering with images’ metadata. It creates a hash, or unique identifier, of each metadata file that makes it easy to spot edits. The technology can also detect when hackers attempt to replace the entire metadata file rather than edit it.
Anthropic plans to release tooling that will make it easier for users to detect Claude-generated content. However, the company cautioned that the technology won’t be perfect. In particular, it may not always detect watermarks embedded in short or heavily edited text snippets.
OpenAI Group PBC has also signed the AI Act’s voluntary Code of Practice. The provision went into effect last week, which means that the ChatGPT developer may soon launch its own text watermarking mechanism. OpenAI created an implementation of such a feature in 2024 and sits on the steering committee of the consortium that develops C2PA.
Image: Anthropic
Support our mission to keep content open and free by engaging with theCUBE community. Join theCUBE’s Alumni Trust Network, where technology leaders connect, share intelligence and create opportunities.
15M+ viewers of theCUBE videos, powering conversations across AI, cloud, cybersecurity and more** 11.4k+ theCUBE alumni**— Connect with more than 11,400 tech and business leaders shaping the future through a unique trusted-based network.
About SiliconANGLE Media
theCUBE AIand theCUBE SuperStudios — with flagship locations in Silicon Valley and the New York Stock Exchange — SiliconANGLE Media operates at the intersection of media, technology and AI.
Founded by tech visionaries John Furrier and Dave Vellante, SiliconANGLE Media has built a dynamic ecosystem of industry-leading digital media brands that reach 15+ million elite tech professionals. Our new proprietary theCUBE AI Video Cloud is breaking ground in audience interaction, leveraging theCUBEai.com neural network to help technology companies make data-driven decisions and stay at the forefront of industry conversations.