Post

It’s unacceptable for a tool to sacrifice an iota of clarity, coherence, meaning, quality, etc. for the purpose of embedding hidden clues within the text to suggest its provenance. daringfireball.net/2026/08/anthro…

Suppose the model has a 49.9% probability of choosing banana, and 50.1% pineapple. Today, without watermarking, you run the model twice and get two different results. Sure, watermarks modify the probability. But you’re still getting a word it was *already* choosing half the time?

Banana, the “inferior” word, was already showing up in your output before. I don’t think watermarking makes the model produce something it wasn’t already doing. I could be misunderstanding how it works though, it’s all kinda way over my head. 😅

I read through John's very long rant, as well as the supporting documentation and referenced essay and... I think this is a lot of ado about nothing. Nothing practical or perceptible, that is.

You will note that Anthropic doesn't provide before / after examples. Nor do they yet provide a watermark detector. Google apparently has been using this in Gemini for a while and don't provide a detector either.

This morning's Dithering episode clarified the argument for me. I do think it's strange they don't provide before & after examples in their documentation. dithering.passport.online/member/episo...