That training data does not preserve provenance seems a "smoking gun" in terms of intent to plagiarize.