But what do you do if your dataset is as inclusive as possible—say, something approximating the entirety of written English, some hundred billion words—and it’s the world itself that’s biased?

Brian Christian, The Alignment Problem, loc. 62375