I'm waiting for tricksters to spread poison-data which causes models to often generate text with banned terms.

Then users complain because "it's stuck with a useless error", as--behind the scenes--everything they do try gets struck down by the keyword-censorship monitor.