You could do it transparently, and without targetting certain content.

Assuming recommendation engines are showing people what they normally watch, you could simply have every X recommendation be whatever your model says is the polar opposite of that. Label it specifically something like "a change of pace", "counter points", or whatever, to make it transparent.

I do wonder if society has got so used to algorithmic bubbles, that just exposing people to the opposite viewpoint would cause outrage.