What's this about?
As of Aug 2, every newly released Claude model embeds machine-readable markers in what it produces: invisible watermarks woven into text, plus signed metadata attached to generated files. It's Anthropic's move to comply with the EU AI Act's Article 50 transparency rules — but it's switched on globally, not just for EU users.
What exactly is a "text watermark"?
A watermark here is a hidden signal woven into the output that a detector can later read to say "this came from an AI," without changing how the text looks to you. For files, it's a digital signature in the metadata. The point is provenance — being able to tell AI-generated content apart from human-written content downstream.
Why is this notable?
It's the first concrete, shipped compliance response to the EU AI Act from a frontier lab, and because it's global, it quietly sets a norm for everyone using Claude. It also feeds the week's provenance theme — knowing what an AI wrote is becoming infrastructure, not an afterthought.