You could do it in different ways, but if you're using synthetic data then you can pick and choose what kind of data you generate which is then used to train these models; that's a way of baking in the censorship.