How do we know AI is ready to be in the wild? Maybe a critic is needed

Can the paradigm of adversarial AI make algorithms safer to send out into the world?