Anthropic Adds Invisible Watermarks to Claude Text Amid EU AI Regulations
Athropic is making changes to its language model, Claude. The company’s adjustments are a response to new regulations from the European Union.
The push for transparency in AI-generated content comes from Article 50 of the EU AI Act. This article requires companies like Anthropic to be more open about their use of this technology. As part of complying with these regulations, Anthropic has signed up to the European Union’s Code of Practice covering AI-generated content.
Starting August 2, 2026, new Claude models will include watermarks at the model level. These watermarks appear in all generated text - regardless of how it’s accessed. This includes major platforms like AWS and Anthropic’s own cloud customers such as Google Cloud and Microsoft Foundry among others.
The watermarks are one aspect of a broader plan by Athropic to increase transparency around AI-generated content. The company is also working on a system for supported files that will include signed provenance metadata based on the C2PA standard.
This framework allows users to record information about digital content, enabling verification whether a file has been processed or altered in any way. For instance, it’s particularly important when verifying AI-generated images like those created using Claude for artistic purposes - which can be used for various applications including advertising and educational materials.
Anthropic cautions that text watermarks nor signed provenance metadata should not be considered definitive proof of authorship. The company notes its models can process original material from people, resulting in the output carrying a mark indicating it was processed by Claude even if not generated entirely by humans.
The technical challenges surrounding these signals are significant with text watermarks able to travel and file metadata potentially disappearing during format conversions or re-saving. Athropic is working on providing more detailed documentation as its marking systems mature.
For businesses that rely heavily on AI tools for data analysis, the addition of watermarks marks a crucial step towards greater transparency in AI-generated material - something that can make a big difference when assessing credibility. By providing machine-readable identity, Anthropic’s watermarks offer another way to verify whether content has been processed by an AI model.
Older models won’t get these new features right away but Athropic is working on adding marking capabilities during the transition period. This will ensure all users can benefit from increased transparency in generated content - something that could help platforms more easily identify when text was processed using Claude.
The changes may also impact how customers use AI-generated images which are used for various purposes including advertising or educational materials. By providing a way to verify whether an image was created with the model, businesses can have greater confidence in their content and avoid potential issues related to mislabeled material.
As new technologies emerge like AI tools that generate data analysis solutions - transparency becomes increasingly important. With watermarks and signed provenance metadata Anthropic is taking steps to address these concerns by providing a way for platforms and organizations more easily identify when a piece of text has been processed using its language model.