The Atlantic scrutinized an AI training data set called Books3, composed of “large, unlabeled blocks of text” and extracted ...