LLM Copyright/Plagiarism filters trivially bypassed with 0% detection [pdf]

I evaluated the Copyleaks copyright/plagiarism detection platform’s ability to detect phonetically and semantically equivalent text that differs orthographically from the original source using tools I made and link to in the paper.

Copyleaks’ own marketing materials cite accuracy rates above 99 % on “paraphrased and disguised” text. In contrast, our trials yielded detection rates as low as 0 %, with multiple transformed works passing undetected despite maintaining near-verbatim semantic and phonetic equivalence to the originals.