When an AI tries to distinguish between a dog and a frog there is no problem. But in certain cases there are limitations in what can be accomplished. In fact some university researcher have found out that a single pixel change can make an AI think that the object is not that type just obtained in the prior test. In fact deep networks are trained and work like a brain. There are neural networks impossible to compute, just think this problem as the cake you baked for example today that is not the same of yesterday but you used the same recipe. The argument goes on as it is fueling an enormous debate among scientists.
Links
https://spectrum.ieee.org/deep-neural-network
