But, according to the paper, that's not what's happening

It's examining published news / research / whatever (input), making statistical predictions, and then comparing (playing) it against other predictions to fine-tune the result