tl;dr: It's a proposal to create a new set of Unicode that LLMs will output with instead of standard output so that their output is watermarked.
It would be fairly easy to undo, but the point is to create provenance for good actors, not fight bad ones.
Also, you can use more robust models (like signing) to detect tampering; this is a markup approach for blended text; whether you trust the sender or not is another matter.