Anthropic shares more details about how Claude’s new watermarks will work

Anthropic has released a blog post detailing how Claude's chatbot will incorporate watermarks to comply with the EU AI Act's Transparency Code. The watermarks aim to make it possible to identify AI-generated content. When Claude selects between words, like "overcast" and "grey" to describe the weather, it creates a pattern in its responses that is detectable only by those with the appropriate key.

However, this watermark does not alter the quality of Claude's output; a reader cannot distinguish a watermarked response from an unwatermarked one. Anthropic plans to release a watermark detection API. Watermarking differs from AI detection methods that look for specific patterns in writing, such as the construction "it's not [X], it's [Y]."

While it may be possible to rewrite text to hide the watermark, significant editing is required, and the text may no longer be considered AI-generated if every word is replaced. The watermark's detectability depends on the text length and the extent of Claude's editing. Light editing by Claude will result in the watermark being detectable in nearly all the words, as the human author will have written most of it.

Code tends to have a weaker watermark because the model must generate working code, limiting the freedom to choose between equally valid options. However, in areas where there's an arbitrary choice between specific words or terms within the code, such as comments, Claude's watermark will still have a negligible impact on the actual code produced.

Claude is not the only AI chatbot set to generate watermarked text; other major model developers have also signed the same Code of Practice and will implement their own watermarks.