A tactic people use to avoid algorithmic censorship: "Voldemorting"

You talk in ways that will be very clear to the listener about what you mean, but that are oblique enough that the algorithm doesn't notice.

Apparently in China people will find characters that look like the word you're trying to say, but don't sound like it.

But communication only works if the receiver can recover the intended signal, so these indirections have to be obvious enough for a receiver to understand.

And particularly effective he-who-must-not-be-named phrases will become used more often (since they effectively communicate the intended message)... thus making them easier for an algorithm to detect and search for.

That means that the replacement phrases must constantly be evolving.