I read the UMD study on why AI text is detectable and built a Claude skill around its main finding: cleaning up vocabulary fixes almost nothing

A study from the University of Maryland and Google DeepMind (arXiv:2604.03136) came out this spring and kills the way most "humanizers" work, including the prompt I'd used for a year. They compared 61,608 texts written by humans and five models (Claude, GPT, Gemini, DeepSeek, Kimi). Then they took the AI texts, ran them through real surface editing — clichés, purple prose, and redundant exposition stripped out by a span-level rewriting framework — and pointed a classifier at narrative structure alone. Detec