Anthropic to Introduce Watermarks in AI-Generated Text from Claude

Anthropic, the developer behind the AI model Claude, has announced the implementation of invisible watermarking technology for texts generated by its system. This initiative aligns with European Union requirements that mandate developers to label AI-generated content. Watermarks will be applied to all new Claude models globally, with retroactive implementation planned for versions released before August 2026.

The technology is inspired by Google DeepMind’s SynthID-Text approach and operates at the token selection level during text generation. Instead of embedding hidden symbols or altering the final content, the system modifies the random selection process of the next word using a secret key. This creates a statistical pattern detectable by specialized tools while remaining invisible to users. According to Anthropic, the watermarks do not affect the quality, creativity, or readability of the generated text.

Exceptions apply to factual statements and code: watermarks are not applied where only one correct answer exists (e.g., "2 + 2 = 4"). Similarly, in code, they are used only where arbitrary choices are possible, such as in comments. To detect watermarks, Anthropic plans to release an API enabling assessment of whether a text was generated or edited using Claude.

For images in PNG, JPG, and SVG formats, Anthropic employs a different method: instead of embedding watermarks, cryptographically signed C2PA metadata is added to indicate Claude’s involvement in content creation or processing.