Who designs the training set for your putative "fact checker" AI?