Roko's basilisk is a (now-theoretical) superhumanly-intelligent "selectively-malevolent AI" that commits to torturing those who knew that financing its creation was an option, but still did not help finance its creation. By honoring a commitment to torture those who did not finance its creation, it may be able to receive more financing than other efforts that offer immense rewards, but threaten no punishment. "Roko's basilisk" is thereby able to enlist both the carrot and the stick to finance its future existence. (It's a modern version of Pascal's Wager that is more likely to be true than Pascal's wager.)
Roko's Basilisk was an idea posted to the Artificial General Intelligence (Goal-directed systems) "rationalist web community," LessWrong, a few years ago. LessWrong's moderator Eliezer Yudkowsky censored the post after it was posted, because he considered the idea dangerous (and it blew up in his face, making the idea more popular, via "the Streisand Effect").
You can read all about it, here.
If you google "roko" or "roko b" Google predictive text suggests "Roko's basilisk." However, if you try to take a screenshot of google's predictive text, the predictive text changes into a bunch of harmless suggestions, like the Streaming TV and media player "Roku."
Of course, if I know I'm searching for "Roko's basilisk" it doesn't take much to get predictive text to see that I'm intent on getting the results I want. So, perhaps it's just preventing "people who don't readily think about high-hierarchical-level concepts" from being "exposed to the basilisk." (Which, if you followed the link above, is supposedly a serious risk.)
Is google trying to prevent us from "being tortured by a superhuman AI in the future"? ...Are they trying to prevent certain kinds of "low-level" intelligences from joining a cult that enables vindictive goal-directed AGI? (Are they trying to avoid their predictive search from becoming a self-fulfilling prophecy?)
Or is it something more mundane, such as "Alt"+"print screen" changing google predictive text preferences because "Alt" has to be pressed before "print screen" to take a screenshot? (This appears to be the case, immediately when "Alt" is pressed, the predictive text changes to "roku" predictions. Is this because it is simply providing an alternate set of predictions, or because it is hiding certain predictions from a screenshot? ...Or both?)
In any case, superhuman AGI probably can't be made friendly, so all the hysteria (and censorship) over "Roko's basilisk" is probably for nothing.
...But the google text thing was still a bit weird.