Anthropic to Embed Invisible Watermarks in Claude-Generated Text
Anthropic announced that new Claude models released on or after August 2nd will embed an imperceptible watermark directly into generated text. The company also plans to provide detection tools to third parties and is working to backport this feature to older models. This move aligns with global regulations like the EU AI Act and helps industries like publishing and education combat plagiarism and verify content authenticity. It marks Anthropic as the second major AI lab, after Google DeepMind, to implement text watermarking at scale. While the watermark survives copying, pasting, and minor edits, it can be bypassed by heavy rewriting, translation, or mixing with other text. Additionally, a detected watermark only indicates Claude was used, which could mean simple proofreading rather than full authorship.
## BACKGROUND
Text watermarking for Large Language Models (LLMs) typically involves mathematically embedding hidden patterns into the token selection process during text generation. This allows detection algorithms to identify AI-generated text without affecting its readability or meaning for human readers.