Nothing you generate with Claude today is watermarked, and nobody can check for marks anyway. What I found after actually reading it all

Honestly, my first reaction to the announcement was mild panic. I reread it three times before it clicked: the marking applies to models launched on or after August 2, and every model you can pick today came out earlier. So nothing you generate right now carries a mark. There is also nothing to check with, the detection API is announced and does not exist. Half the threads here missed both facts. Then I dug further and found the part that did make me angry, and it has little to do with the marking itself. In a nutshell the mark cannot tell a text the model wrote from your own text (say translating), generated from scratch or the model fixed "heavily edited this" (heavily means substituted a few synonyms). Anthropic says this straight in their FAQ. And whoever eventually points a detector at your writing will not spend a minute on that difference, you will just be "flagged as AI". I write in one language and publish in another, so this one is personal: translate your fully human text with a marking model, and statistically it becomes 100% machine-picked words. The law that started all this simply has not understood the topic yet, and I think that difference, written text versus touched text, is the whole conversation we should be having. I went through the docs, the papers and these threads and wrote it all up in plain words, with a table of who actually marks text today. Link in the comments.