300 local LLM rewrites vs. AI text watermarks

On October 3, 2026, a write-up on DEV Community described a hands-on test of whether AI text watermarks can be stripped by paraphrasing. The author ran 300 rewrites on a single GPU with 16 GB of VRAM, passing text through another model with the instruction to rewrite it in its own words.

The premise, as the author lays it out, is that AI text watermarks hide a signal in which words a model picks. If the signal lives in word choice, a natural idea is to run the text through a second model and let it rephrase, breaking those patterns. That raises the question the test was built around: if you rewrite a text enough to break its word patterns, do you still keep its meaning?

The short answer, according to the post, is yes, mostly. Rewriting largely removed the watermark while preserving the original meaning. But the author reports that the model you pick matters a great deal, and that there is a real trade-off between how much the text changes and how much of its meaning survives. Push the rewriting hard enough to be safe and the meaning starts to suffer; keep the meaning intact and the watermark may survive.

The post does not name the models used, the watermarking scheme tested, or the metrics behind the 300 runs, and it gives no detail on how meaning preservation was judged. Its account stops at the two conclusions: rewriting works in most cases, and model choice together with rewrite strength are what decide the outcome.