Unicode merges some Chinese and Japanese characters into same code points and leave others as they are, on random per-character basis. So Chinese guys gets a ton of Japanese mixed texts and it causes tons of problems.

Japanese users are angry when system chooses Chinese fonts over Japanese, Chinese users apparently suffer with a font mishmash of Chinese and Japanese all the time like a mid-word capitalization. The official sanctioned solution is to just commit a genocide and nuke the offending language out of the system or to attach IVS to every Japanese characters which is like zero padding every single characters by a byte or two.

It's really putting everyone in pain and making absolutely no one happy, which sounds like what a good compromise tends to be, but LLMs seem to be struggling with there being multiple completely separate language families sharing codepoints but not character shapes or syntax. I've seen a smaller LLM model use made up verbs spilling over from the other language. I believe ML guys don't like datasets that do that kind of things. Same apply to image models.