LLM watermark reversing

[chants]

The repo is a tragedy in terms of removing text watermarks.

"Disclaimer: what removing a text watermark costs
Text watermarks live in the wording itself: the signal is spread across token choices, so nearly every sentence carries a little of it. Two consequences follow, and they are why Layer B is honestly described as best-effort rather than a magic eraser.

Removal means rewording, not restructuring. Shuffling paragraphs, changing headings, or light touch-ups barely move the signal. Stripping a statistical mark requires rewriting a substantial fraction of the text — sentence by sentence, not section by section.

Rewording degrades the copy. Any rewrite replaces the original word choices with the rewriting model's, which flattens tone, voice, and precision. On production copy (SEO, marketing, client work) that degradation is real and often visible to the people who care most about the writing. It is like taking text from a top-tier model and asking a less capable model to rewrite it from scratch: the result cannot exceed the rewrite model's ceiling."

Rewording was not my suggested approach. So I see this a weak inferior method due to lack of creativity. Got there first with a weak idea. Any tool needs to be dual use and quiet. As companies could start blocking or taking action on user accounts, and even have no choice due to regulation. This is a one hour easy job, so id be wary of using that repo for anything mor than watermark detection.

[deepzero]

The statistical token choice nudge just has to encode something (probably your account id soon...), they can encode it however they want. So sentence based rewording is likely really the only chance. Even then it's difficult.

Probably soon the first skids exploit code will be linked to his claude account by sampling the commands used... dark times...