True although in the EU at least I'm pretty sure that retention of the effect of some data after aggregation is ok as long as the originating data is deleted an

True although in the EU at least I'm pretty sure that retention of the effect of some data after aggregation is ok as long as the originating data is deleted and the aggregation is such that the original data couldn't be reverse-engineered afterwards. So that would cover use of your data in an ML training set. They would need to remove it from the set for future trainings but the weights of existing models wouldn't need to be altered. That's my understanding anyway.