I would imagine that the difficulty lies in finding effective ways to remove information from the training data in that way. There's an enormous amount of data, and LLMs are probably pretty good at putting information together from different sources.