AI Signal 439
Anthropic adds invisible watermarks to Claude outputs for EU AI Act compliance
↗
Illustration only Photo by Anne Nygård on Unsplash
Anthropic implemented invisible, machine-readable watermarks on text generated by supported Claude models across all products and APIs worldwide to comply with the EU AI Act's Article 50(2).
The watermarks survive copy-pasting and light editing but can disappear through substantial rewriting, making them a hint of provenance rather than proof of authorship. This blanket implementation may disproportionately burden users who rely on AI for legitimate assistive purposes while being weakest against motivated deception.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Anthropic applied invisible watermarks at the model level across all supported Claude products and API surfaces worldwide.
The watermarks survive light editing but can disappear through substantial rewriting, paraphrasing, translation, or mixing.
Anthropic's implementation goes beyond the EU AI Act's minimum requirements, which exempt systems performing assistive editing functions.
THE CLUSTER