As an anecdote, a few days ago Kagi's Research (Experimental) model had no problem generating critical images of Trump and Vance. But yesterday it expressly refused to generate the likeness of Pam Bondi.
(This is surely the underlying component models' censuring, not Kagi's.)
That’s anecdotal evidence demonstrating the wonders of simple probability. There’s so much opacity to refusals that claims like this (currently) can’t have enough rigor to be used to make an argument.
To strongly make the claim you’re making you’d have had to perform extensive before/after tests to be able to compare.
That said - the lack of transparency that’s making you feel uncomfortable with this whole thing, that is concerning.
(This is surely the underlying component models' censuring, not Kagi's.)