They're probably also being trained on the contents of preprint services like arxiv, which includes lots of papers with little or no merit.