Anthropic explains how Claude's new text watermarks will work

Anthropic has published a blog post answering key questions about the watermarking system it plans to add to text generated by its Claude chatbot. The move, announced earlier this week, is meant to comply with the EU AI Act's Transparency Code, which requires AI companies to make AI-generated content identifiable.

The company explained that Claude will create patterns in its responses when making low-stakes choices, such as picking between similar words. These patterns are invisible to readers but can be detected by anyone with the right key. Anthropic said watermarking won't affect the quality of Claude's output, and that a watermarked response will look the same as an unwatermarked one.

Anthropic will use the SynthID-Text approach, first outlined by Google DeepMind in 2024, and plans to release a watermark detection API. The company noted that this is different from AI-detection tools that look for writing patterns, since watermarking relies on a hidden code rather than stylistic tells.

The company acknowledged that a complete rewrite of the text could remove the watermark, but light editing probably won't. For text that Claude only proofreads or lightly edits, there may be very little for the watermark to attach to, since most words would come from the human author.

When it comes to code, watermarking will have a smaller presence, because the model must produce working code and has less freedom to choose between equally valid options. However, watermarks could still appear in areas like comments within the code.

Anthropic also said Claude won't be the only chatbot with watermarks, as other major model developers have signed the same Code of Practice and will be implementing their own systems. Meanwhile, some users have reacted negatively, with reports of subscription cancellations and online debates about the move.

Source: TechCrunch