Hacker Newsnew | past | comments | ask | show | jobs | submitlogin
Anthropic Posts 'How Claude Marks AI-Generated Content' Without Explaining How (daringfireball.net)
6 points by nozzlegear 32 days ago | hide | past | favorite | 7 comments


it weaves an imperceptible watermark directly into the text itself. You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response

That's John Gruber quoting Anthropic. Gruber's completely correct on this one. Either there's non printable Unicode characters, or the word choice is biased, and the Anthropic statement is lies.

Does Anthropic really think they can claim magic and everyone will just accept it, or have they spent too long communing with Mr Flattery and can no longer distinguish reality from fantasy?


Anthropic is saying the word choice is biased without affecting readability and meaning.


That still seems trivial to strip out and/or launder. Seems like theater for those who are immaterially concerned (in this case, the EU).


It seems we generally have trouble removing emdashes and phrases like "load bearing", even when we're looking for them. Identifying and removing subtle contractions, or less visible nuances will be tough, especially if they change to avoid detection.


prove it


I don't see how that's possible especially for shorter blocks of text.


are you a cryptographer?




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: