Anthropic Clarifies Claude Watermarking for EU AI Act Compliance

+AI Analysis - Key Points
- 01
Anthropic will use the SynthID-Text approach to watermark Claude's output
- 02
Watermarking does not affect output quality, but can be detected with a key
- 03
Light editing may not completely remove the watermark, but a full rewrite will
Anthropic has sought to address concerns about its plan to watermark the text generated by its chatbot Claude. The company published a blog post on Friday, explaining how the watermarking process will work and how it will impact the quality of Claude's output.
The move is aimed at complying with the EU AI Act's Transparency Code, which requires AI companies to use systems that make it possible to identify AI-generated content. This has sparked debate among Claude users, with some expressing concerns that the watermarking could be used to hide the fact that the content was generated by AI.

In its blog post, Anthropic explained that the watermarking process will involve creating a pattern in the text that is undetectable to the reader, but can be detected by anyone with a key that encodes it. The company emphasized that watermarking does not impact the quality of Claude's output, and that a watermarked response is indistinguishable from an unwatermarked one.
More specifically, Anthropic said it will be using the SynthID-Text approach, which was outlined by the Google DeepMind team in 2024. The company also plans to release a watermark detection API, which will allow developers to detect the watermark in text output.
The company acknowledged that it is possible to rewrite the text to hide the watermark, but noted that light editing may not completely remove the watermark. A full rewrite, where every word is replaced, would be required to remove the watermark completely.
In this case, it's arguable whether the text can still be described as AI-generated, as the watermark would be removed. However, Anthropic emphasized that the watermarking process is distinct from AI detection approaches offered by companies like Pangram, which look for 'tells' in the writing to reveal AI usage.
The company also noted that the detectability of the watermark will depend on the length of the text and how heavily Claude has edited it. If the text has only been lightly edited, nearly all the words will have been written by the human author, and the watermark may not be detectable.
Overall, Anthropic's clarification on the watermarking process aims to address concerns among Claude users and provide transparency on how the company plans to comply with the EU AI Act's Transparency Code.
VideoVideo Coverage
Anthropic Clarifies Claude Watermarking for EU AI Act Compliance
Watch video coverage on YouTube
Opens relevant video results on YouTube.