Claude's next models will start tagging their own text and images as AI-generated, and that labeling is headed to every Claude user, not just Europeans.
Anthropic published guidance explaining how new Claude models will comply with Article 50 of the EU's Artificial Intelligence Act, which set an August 2, 2026 deadline for identifiability measures. Text will be watermarked by biasing token selection during generation, leaving a statistical pattern that detection tools can later flag; passages under 200 tokens are exempt because they don't carry enough data to mark reliably. Images in SVG, PNG, and JPG formats get a C2PA provenance certificate that breaks if the file is altered, though Anthropic notably stops short of committing to actual pixel-level image watermarking, despite the EU's Code of Practice calling for it. Existing models won't get any of this until a second deadline, December 2, 2026.
Here's the part the EU framing buries: Anthropic's guidance opens by describing "models launched in the EU," but later paragraphs clarify that marking will apply to output from supported models wherever Claude is offered, worldwide. A compliance deadline aimed at one regulator is quietly becoming a global product change, meaning US and other non-EU users get labeled outputs whether their local law demands it or not.
OpenAI and Google shipped similar disclosures months ago, so Anthropic is playing catch-up here, and its own guidance concedes that a missing watermark is no proof content wasn't AI-generated.