Anthropic will embed imperceptible watermarks in text and files generated by its Claude models, starting with those launched in the EU. The move aims to comply with the EU's AI Act, but may face skepticism from users.
TL;DR
- Anthropic introduces watermarks in Claude's AI outputs to comply with EU transparency rules.
- Watermarks will be applied globally, affecting all Claude products and third-party providers.
- Users and researchers express doubts about the effectiveness and permanence of text-based watermarking.
What happened
Anthropic announced it will embed watermarks in text and files generated by its Claude models, starting with those launched in the EU. This effort is part of its compliance with the content and transparency rules in the EU's AI Act. The company plans to add output marking to already released models during the transition period allowed under EU law.
The watermarks will be applied to output from supported Claude models on various platforms, including the Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag. This also includes third-party providers of Anthropic models like AWS, Google Cloud, and Microsoft Foundry.
Anthropic stated that the watermarks will be imperceptible and will not change the meaning, quality, or readability of Claude's responses. The watermarks will travel with the text when copied and pasted, and may persist through some editing. For files, Anthropic will rely on signed provenance metadata that conforms to the C2PA standard.
Why it matters
This move by Anthropic may amplify the appeal of open weight models, as users might seek alternatives that do not embed watermarks. It could also alienate Claude customers who do not want the provenance of their AI-generated content to be known.
Users and researchers have expressed skepticism about the effectiveness of text-based watermarking. Researchers have already demonstrated that image-based watermarking can be undone, and it may not be difficult to create an optical character recognition system that strips or omits obscure marks from Claude-generated text.
Anthropic itself acknowledges that detected marks are not conclusive evidence that Claude produced the content, and the absence of marks cannot guarantee that AI wasn't involved in the creation of a particular piece of content. However, the scheme may be sufficient for legal compliance.
Key facts
- Anthropic will embed watermarks in text and files generated by its Claude models, starting with those launched in the EU.
- Watermarks will be applied globally, affecting all Claude products and third-party providers like AWS, Google Cloud, and Microsoft Foundry.
- The watermarks will be imperceptible and will not change the meaning, quality, or readability of Claude's responses.
- Anthropic will rely on signed provenance metadata that conforms to the C2PA standard for files.
- Users and researchers have expressed doubts about the effectiveness and permanence of text-based watermarking.
- Anthropic acknowledges that detected marks are not conclusive evidence of Claude's involvement in content creation.
- The move may amplify the appeal of open weight models and alienate Claude customers.
- Anthropic plans to provide details about how people can detect Claude's marks in forthcoming documentation.
Context
The EU's AI Act aims to ensure that AI systems are safe, transparent, and respect users' rights. The act includes rules on content and transparency, which require AI systems to disclose when content is AI-generated.
Anthropic's move to embed watermarks in AI outputs is part of a broader trend among AI companies to comply with regulatory requirements and address concerns about the provenance of AI-generated content.
The effectiveness of watermarking techniques remains a topic of debate among researchers and users. While watermarks can help trace the origin of AI-generated content, they can also be circumvented or removed, raising questions about their reliability.
