Claude Watermarks Its Content

Since August 2, 2026, Anthropic has embedded a machine-readable mark into everything Claude generates or edits. For text it's an invisible watermark right in the output that survives copy-paste; for .png, .jpg and .svg files it's provenance metadata following the C2PA standard. It works globally, not just in the EU — the company's way of meeting the EU AI Act.

What matters is what the mark does NOT prove. It only signals that Claude "may have processed" the content — not that it wrote it: the mark stays even when the model merely proofread, translated or summarized someone else's text. And it disappears after heavy rewriting, on very short passages, and once a file is re-saved with its metadata stripped. How it works is broken down by the Semrush blog.

What content teams should do now:

• Set a baseline: how AI content ranks and gets cited in ChatGPT, Perplexity, Google AI Mode — so you can see the effect before anything changes;
• Don't treat the mark as a verdict: its presence ≠ "fully AI-written";
• Write down an editorial policy: where AI helps and where manual work and fact-checking are required;
• Watch the platforms: search engines haven't announced adoption, but the infrastructure now exists.

The practical question has shifted from "can you tell AI text apart" to "what will platforms choose to do with it." Prepare for a scenario where the mark affects ranking or trust — and build a process where an editor gives the content its value, not the fact of where it came from.