Yeah that was actually what I was trying to get at. Microsoft would have to get rid of the model trained on protected content, remove that content from the training set, and start over training a new model.