AIBRIEF
Anthropic's Claude became the first major AI assistant to embed invisible watermarks directly into the AI model itself to comply with Article 50 of the EU AI Act. Other leading AI companies selling in Europe now face the same rule.
Anthropic on Aug. 2 shipped model-layer watermarks on Claude, embedding invisible markers into text and signed provenance metadata into supported files for new models released in the EU on or after that date. The change responds to Article 50 of the EU AI Act, the bloc's binding AI transparency rule that became legally applicable the same day.
The watermarks live at the model level, not the app layer, so the marks persist regardless of which Claude interface generates the text. Anthropic says they can survive copying, pasting, and some editing. Provenance metadata functions as a signed receipt inside supported files, telling downstream tools that the content was AI-generated. Separate rules under Article 50 also require disclosure labels on certain deepfakes and synthetic public-interest images and video.
Anthropic has also signed the EU's voluntary Code of Practice on Transparency of AI-Generated Content. The underlying obligations apply whether or not a provider signs, and they now cover every frontier lab selling into the EU. The remaining question is execution.
That test is live. Watermarks that don't survive screenshots, heavy editing, or translation become a tax everyone pays; the same is true if the layer degrades output quality or latency. Anthropic has a 90-day head start to show whether model-layer compliance is a moat or a baseline the rest of the industry catches up to.