I'm not aware of Apple doing the kind of modelling you're talking about. Instead, our machine learning tends to be on device intelligence.
For instance, as the keyboard learns about the user, the device itself collects and uses this information to train itself for that user without the information leaving the device. Where we are collecting data to help inform community models, we're using things like local differential privacy, which applies randomization to the data before it leaves the user's device, so we're not able to go back and tie the individual user inputs and their content to a user. It's very much a device focus for us.