Maybe they have an AI that ingests thinkpieces and twitter feeds lambasting them for insufficient censorship, and based on the strength of sentiment and number of retweets apply additional censorship to the specific examples listed, with the cost function being something like f(likelihood_of_negative_public_opinion) + g(impact_on_search_quality).

If I were to do this I'd allow penalized results to resurface from time to time, and leave it up if nobody tweets too hard.