XNews

Stratechery: Claude Watermarking Spills EU Rules Globally, Undermining Human Authorship

Anthropic added model-level watermarking to Claude outputs to comply with EU AI transparency rules, but the change applies globally. The watermark likely uses a green/red token list scheme (arXiv 2301.10226) or Google's SynthID tournament sampling. Detection is not definitive: editing and translation can add or remove marks. EU rules treat AI as a separate entity, but the author argues AI is a tool; requiring watermarks credits the tool over the human. The rule forces creation and realization to be bundled, contradicting the unbundling trend where AI breaks the last bottleneck of idea-to-output. The author concludes the EU is misguided.

Source posts