- Anthropic states that new Claude models will automatically watermark text to comply with the European Union AI Act.
- The watermark adds no characters, does not change the content, and does not affect quality, readability, or usage costs.
- The technology is deployed globally as Anthropic cannot yet limit it by specific regions.
- Claude utilizes the SynthID-Text method published by Google DeepMind in the journal Nature in 2024.
- The mechanism operates by selecting between words with equivalent meanings through a special random source instead of a standard random number generator.
- Readers cannot perceive the difference, but a system with a decryption key can estimate the probability that the text was generated by Claude.
- The watermark does not contain information identifying individuals, organizations, or conversations and cannot be traced back to the user.
- Detection effectiveness increases with text length; short snippets or content with very few word choices are harder to detect.
- When Claude only performs light editing on human-written text, the amount of change is often too small to create a clear identification mark.
- Source code has a lower level of watermarking because many segments require absolute precision where only one choice is correct.
- Claude will provide a detection API in the near future to check the probability of text being created by Claude.
- For images and files like PNG, JPG, or SVG, Claude does not use watermarks but embeds metadata according to C2PA standards to record content provenance.
- Anthropic notes that minor edits usually do not completely remove the watermark, but rewriting the entire text may erase the sign.
📌 Anthropic states that text watermarking is not intended to identify users but only helps assess the probability that Claude participated in creating or editing the content. This technology operates silently, does not affect the reading experience, and is based on Google DeepMind’s SynthID-Text method. It is part of the effort to meet the transparency requirements of the European AI Act, while paving the way for the identification of AI content on a large scale in the future.
