Roger Montti dug in with some research to try and understand how Claude’s watermark might work by looking at academic papers and related materials. My take-away is that it is not hidden images, text, or such, but rather the patterns and use of letters, words, and phrases somehow being a reflection of AI token usage. Hmm, “token usage” maybe being related to how AI selects words to use? In any event, it seems that it can largely survive editing and summation.

How AI Watermarking Works
Roger Montti @ searchenginejournal.comhow • 3 days ago








