If this sort of error isn’t acceptable, it should be part of an evaluation set for your discriminator

Fundamentally I’m not disagreeing with the article, but also think most people who care take the above approach because if you do care you read samples, find the issues, and patch them to hill climb better