Re:Anthropic's 'Watermark' Text Adulteration in Claude Is a ...
by kfse on Monday August 17, 2026 @10:26AM
(#49331565)
The model doesn't particularly "know" which tokens within its output are "comments", it's not so easy
Re:Anthropic's 'Watermark' Text Adulteration in Claude Is a ...
by alasdair_ on Monday August 17, 2026 @02:05PM
(#49335150)
It knows when several tokens are about equal vs. times where one token is vastly preferred. In the latter case, that’s usually code or math or something similar and so it won’t alter those tokens.
That's a fine place to put the watermark to track those users who accept the code blindly and don't delete/edit the comments.