- Anthropic's latest Claude models are embedding invisible, machine-readable watermarks into generated text—marks that can survive copy-pasting and some editing.
- Anthropic releases more information about Claude AI watermarking amid user confusion regarding its implications.
- Claude will leave a hidden trail in AI-generated text, which can indicate that Claude was likely involved in generating a piece of text.
- The watermarking system relies on Claude leaving a pattern in its responses, detectable to any party with a key that encodes it.
- Anthropic's watermarking is designed to comply with the EU AI Act, with several major AI model providers signing the EU Code of Practice on Transparency of AI-Generated Content.
- The watermark cannot be used to trace back to a specific individual, organisation, or chat session where the text was generated.
- Anthropic acknowledged that watermarking is not impossible to defeat, as light editing may not remove it completely.
- Watermarking will not affect the quality, creativity, or readability of Claude’s content.
Anthropic's Claude models are embedding invisible, machine-readable watermarks into AI-generated text, a move aimed at complying with the EU AI Act while addressing writer concerns. These watermarks can survive copy-pasting and some editing, indicating Claude's involvement without confirming full authorship.
The watermarking system relies on Claude leaving a detectable pattern in its responses, which can be identified by parties with a specific key. Anthropic emphasized that the watermarking process will not affect the quality, creativity, or readability of the content produced by Claude. “A watermark can only determine that Claude was likely involved with the content at some point,” the company stated, clarifying that it cannot distinguish between full authorship and heavy editing.4
The company noted that the watermark is less active in passages with singular correct answers, such as code, which is less heavily watermarked due to its need for exact terms. “The more extensively Claude rewrites a piece, the more opportunities there are for the watermark to attach,” Anthropic explained.
This initiative aligns with the EU's push for transparency in AI-generated content, as several major AI model providers signed the EU Code of Practice on Transparency of AI-Generated Content in July. Over the coming months, Anthropic plans to extend watermarking to its older models as well.
The watermarking system is based on the SynthID-Text approach developed by Google DeepMind, ensuring that the process remains invisible to readers while maintaining the integrity of the text.
“The watermark, based on Google DeepMind's SynthID-Text approach, creates a statistical pattern via Claude's word choices without adding hidden characters. It cannot trace content to a specific user, and light editing may not remove it, but rewriting every word can.”
