It all boils down to your ground "truth"
This is a problem for any machine learning system: if your ground truth contains errors, the machine may well learn to copy those mistakes. In the case of deep learning, this problem is compounded, because they require a tonne of data to train. That makes curating the ground truth you feed it very, very difficult indeed. In deep learning you may not need to painstakingly design your features, but what you gain there you pay for in terms of the work needed on getting the ground truth right. There no such thing as a free lunch.