They said it’d even work if the text was substantially edited by the user, though. Maybe there’s multiple layers of redundancy so it just degrades the watermark but it takes a lot of editing to completely dilute it.
I think the first person was right. It’s probably just a pattern based on this part:
It may also survive certain edits. However, the company warned that extensive rewriting, paraphrasing, translation, or mixing the content with human-written material could weaken or remove the detectable signal.
Pattern matching, maybe an extra space every 5 word, ensure the 3rd sentence starts with a T and ends with a P, etc.
They said it’d even work if the text was substantially edited by the user, though. Maybe there’s multiple layers of redundancy so it just degrades the watermark but it takes a lot of editing to completely dilute it.
I can’t think of any pattern that is supposed to survive edits that wouldn’t create a ton of false positives in existing works.
So an average AI and plagiarism detector that already exists.
I think the first person was right. It’s probably just a pattern based on this part: