For me it is all about the training and re-training of the AI.  Once it has been properly trained and believed to be good, there needs to be random checks on those claims that have been rejected.
Just like humans, AI can make mistakes and we need to keep honing their behavior.