This is because the user directly ties the device to their account and they have full visibility into the utterances. You can see a full list of what you've said and you can delete any one of those. Those will immediately get removed from the database and would not be the basis for a recommendation.
On May 29th, 2019. See this statement in context.