So that people don't feed proprietary data into a model that will train itself from it. Standard DLP practice.