I guess I would have expected the dataset to have only the occasional horrible prompt if it were just user-provided data, but I got the sense from the article that these guys were labelling horrific text all day.