In general models trained on individual users' data get rebuilt when the underlying data changes so as to properly reflect the current state of the source data.
Do you have any specific policies around this? I suspect it would be hard to say definitively that there aren't remnants of data I've asked Google to forget still kicking around.