Anthropic has taken a significant step towards complying with the transparency rules set out in the European Union’s Artificial Intelligence Act. The company has begun placing invisible watermarks on texts created or edited by Claude models.

This new system uses a model-level marking procedure to identify the origin of texts. The application was officially launched on August 2nd.

How Does Invisible Watermark Technology Work?

This watermark system developed by Anthropic can maintain its presence even when texts are copied, pasted, or modified in various ways. The company states that these marks are added not only to generated directly content but also to texts written by real people that are edited or adapted using Claude.

This technology is designed to cover all platforms where Claude models are used. Tools like Claude, Claude Code, Cowork, and Claude Tag, as well as systems hosted on AWS, Google Cloud, and Microsoft Foundry, are included in this new transparency rule. While Anthropic hasn’t fully disclosed the technical details of the watermarks yet, it appears the system is designed to digitally track text.

Scope to Continue Expanding

The company states that this new approach is initially limited to new models but will eventually expand to content produced using older models. This strategy aligns directly with the European Union’s legal regulations aimed at increasing transparency in digital content. Users should be aware that these invisible markers are now included in their content when creating or editing text through Claude.

This method, implemented at the model level, has the potential to create a new standard for platforms seeking to verify the source of content. With this move, Anthropic is both fulfilling its legal responsibilities and establishing a more secure system for tracking the origin of text-based content.

The company’s approach to content accuracy and transparency with this technology could set an example for other players in the industry. Further improvements and adaptations to different content types are expected with future updates to the Claude models.

How do you evaluate Anthropic’s move and the invisible watermarks added to the texts?