So that includes all online and offline + offsite backup systems, presumably? And hopefully any such data is "de-trained" from all applicable ml models and systems, of course.
Without going into too many details, yes. The trick is that you can key sensitive data, encrypt it, and delete it by throwing away the key.
I don't know if Google can do de-training (depends on how the training data is generated), but generally if the trained data can't be tagged for removal it also can't be reversed from the output of the training.
I don't know if Google can do de-training (depends on how the training data is generated), but generally if the trained data can't be tagged for removal it also can't be reversed from the output of the training.