How Claude's Watermark May Work

The MirrorMark paper from George Mason University appears to explain how Claude's watermark might work. Anthropic hasn't disclosed the method, but it named six of its properties — and the 2026 research matches them almost word for word. Search Engine Journal reports on the connection.

Claude's mark is embedded during token generation, doesn't change the meaning or readability of the text, survives some editing and paraphrasing, and can be verified by third parties. MirrorMark does exactly that in three steps: it "mirrors" the random token sampling to encode bits without distorting word choice; it places the marks by context through a dedicated scheduler; and the decoder replays the process and declares text watermarked if the signal exceeds a threshold.

What the resilience tests showed:

• Against paraphrasing, the method kept "strong separability" between watermarked and ordinary text;
• Detection accuracy sat around 57.8% at a 1% false-positive rate — noticeable, but not perfect;
• Heavy rewriting degrades it — short and paraphrased passages slip out of detection.

The takeaway for the market: detecting AI text after light edits stays workable but is a threshold call — you can't build hard penalties on it without room for error. For SEO the signal is simple: don't hide where the text came from, own its quality — an editor, not an attempt to fool the detector, is what makes content resilient to any mark.