The success of this hinges in ai training companies converting these human em dashes back to regular em dashes when adding documents to their training corpus.

And those using LLMs from not post-processing the output to swap such known watermarks. Not sure if meant as a joke RFC though.