Anthropic provides technical details on Claude's text watermarking system

Anthropic has released detailed explanations about how its watermarking system for Claude-generated text will function. The company uses the SynthID-Text approach developed by Google DeepMind, embedding imperceptible patterns in text that are undetectable to readers but identifiable with the proper decryption key. The watermarking makes subtle choices in word selection without impacting text quality or readability. Responding to technical concerns, Anthropic confirmed it will release a free API enabling users and third parties to independently detect the watermarks. The company also clarified that light editing will not remove watermarks, though complete rewrites replacing every word will eliminate them. For lightly edited content, the watermark may persist depending on text length and the extent of Claude's modifications. Code presents a different case, as the watermark has minimal presence where the model must choose between functionally identical options. Comments within code may retain watermarks, but this has negligible impact on actual code output. Anthropic noted that other major AI developers have committed to implementing their own watermarking systems under the same Code of Practice to comply with EU AI Act requirements, though the company is among the first to provide public technical specifications.