Anthropic to Implement AI Content Watermarking Under New EU Legislation

Anthropic to Implement AI Content Watermarking Under New EU Legislation

The startup has signed a Code of Conduct for Transparency in AI Content as part of the European AI Act. This came into effect on August 2, and from now on, providers and deployers of AI systems entering the EU are required to embed watermarks in generated text and files.

Anthropic has just updated its documentation on this matter: documentation.

Although this is an EU law, the watermarking will be applied globally. However, this does not mean that from today all your text from Claude models will be marked.

Key Points:

1. Watermark Support:

   • Only models released after August 2 are required to support watermarking from launch.
   • For older models, the law provides a transition period of 4 months, and Anthropic's older models currently remain without watermarks.

2. Watermark Accessibility:

   • So far, no one except Anthropic can read the watermark. There is no public tool available to check text for a watermark.

3. Text Editing:

   • Watermarks disappear if the text is slightly edited. This is related to how the watermark is embedded in the text during the generation process.

Technically, this is interesting, legally mandatory, but as a solution to the original problem – quite leaky.