Anthropic has published a blog post explaining the technical details and implications of its planned text watermarking system for its AI chatbot, Claude. The move, announced earlier this week, is designed to comply with the EU AI Act’s Transparency Code, which mandates that AI companies implement systems to identify AI-generated content. The new post addresses key questions about how the watermarking will work, its limitations, and its impact on different types of text, including code.

The company explains that watermarking involves Claude making subtle, undetectable choices between equally valid words—such as “overcast” versus “grey”—to create a pattern that can be identified by anyone with the appropriate key. Anthropic emphasizes that watermarking does not affect the quality of Claude’s output, stating that a watermarked response is indistinguishable from an unwatermarked one to a reader.

Anthropic will use the SynthID-Text approach, developed by the Google DeepMind team in 2024, and plans to release a watermark detection API. The company also clarifies that watermarking is different from AI detection tools that look for stylistic patterns in the text, noting that checking for a watermark is fundamentally different from picking up on writing “tells.”

The post addresses the possibility of removing the watermark. Anthropic states that light editing likely won’t eliminate the watermark completely, but a complete rewrite where every word is replaced will. The company adds that in such a case, it’s arguable whether the text can still be considered AI-generated.

For text that has been proofread or lightly edited by Claude, the watermark’s detectability will depend on the length of the text and the extent of editing. If only lightly edited, nearly all words will be written by the human author, leaving little for the watermark to attach to.

Regarding code, Anthropic notes that the watermark will have a negligible effect because the model must produce functional code and has less freedom to choose between equally valid options. However, watermarks can be applied to arbitrary choices within code, such as comments.

Finally, Anthropic indicates that Claude will not be the only AI chatbot with watermarking, as other major model developers have signed the same Code of Practice and will implement their own watermarks.

#Anthropic, #Claude, #AIWatermarking, #EUAIAct, #SynthID